Measuring Changes in Corruption over Time∗†

Maya Dalton1 and Justin Esarey2

1
Pennsylvania State University, Department of Political Science
2
Wake Forest University, Department of Politics and International Affairs

July 10, 2024

Abstract
Prior criticism has argued that country-level corruption indicators are incompara-
ble to each other and over time. These criticisms have been targeted at sources of bias
and noise in the measures as well as conflicting conceptualizations of corruption among
them. We present evidence that these indicators all track a shared component identi-
fiable as a within-country change in overall corruption, but despite this, they remain
highly inconsistent with one another. Many causal inference research designs, such as
difference-in-difference and dynamic panel instrumental variable models, rely on such
within-country changes to identify causal relationships. As a result, we argue that
empirical findings about corruption based on these designs will be particularly sensi-
tive to the choice of measure. Most importantly, we present a new synthetic measure
of within-country change in corruption designed to isolate the shared component that
influences existing indicators. Our new measure is robust to bias and correlated mea-
surement error that may be present in those indicators and permits inferences about
the aspects of within-country corruption on which they agree.

∗
We gratefully acknowledge the financial support for this project provided by the Wake Forest University
Undergraduate Research and Creative Activities Center (URECA) and the WFU Department of Politics and
International Affairs.
†
Competing interests: The authors declare none.

1
 Tempora mutantur, nos et mutamur in illis.

Johannes Nas

Introduction
Every measure of corruption is constructed differently, but all face the fundamental challenge
that corruption is a hidden behavior; its perpetrators do not want their largely illegal dealings
recorded in a data set (Brooks et al., 2013, p. 27). Consequently, the “abuse of public power
for private gain” (Hawken and Munck, 2008, pp. 74-75) is difficult to directly observe.1 The
unobservability of corruption creates a large number of persistent and difficult measurement
problems, problems that have been a core concern of scholars and policymakers for nearly as
long as measures of corruption have existed (e.g., Sampford et al., 2006; Brooks et al., 2013;
Heywood and Rose, 2014). A recurring theme in this research is that influential country-
level measures of corruption are not comparable over time. Transparency International’s
Corruption Perception Index (TI CPI, published since 1995) is a frequent target of this
criticism (Andersson and Heywood, 2009; Brooks et al., 2013, p. 37; June et al., 2015, pp. 26-
27), but similar issues have been raised for the International Country Risk Guide (ICRG)
measure of the political risk of corruption as well as the World Bank Governance Indicators
Control of Corruption estimate (WBGI CCE) (Treisman, 2007, p. 220; Knack, 2007; Hawken
and Munck, 2008, p. 85; Standaert, 2015). Criticism has not stopped these measures from
being widely used; the TI CPI alone has been cited about 8,240 times according to Google
Scholar.2
The potential incomparability of corruption measures—to each other, and to themselves
1
While there are alternative conceptualizations of corruption (e.g. Warren, 2004), its definition as “abuse of
public office for private gain” has become relatively standard across the literature since being adopted by
several prominent non-governmental organizations (June et al., 2015, p. 12).
2
We produced this figure by searching with the prompt “Corruption Perceptions Index” and intitle:corruption.
Google Scholar citation counts are estimates and therefore inexact.

1
over time—has been recognized in scholarship and is widely known by experts in the field.
What remains to be done? We believe that a priority of scholarship must be to develop a
country-level measure of the overall pervasiveness of corruption that validly measures changes
in corruption within each country over time. Very basic descriptive inferences require such
measures; without them, we cannot even answer simple questions such as whether democ-
racies have gotten more corrupt (on average) over the last twenty years. Moreover, many
causal inference techniques for observational data (such as difference-in-difference designs or
dynamic panel data instrumental variable models) depend crucially on the accurate measure-
ment of these changes. This concern is already recognized in published work; for example,
Gründler and Potrafke’s (2019, p. 2-3) study of the effect of corruption on economic growth
begins:

Studies using the CPI in panel data models ignored that the CPI was not com-
parable across countries and over time before 2012. In particular, including fixed
period effects in panel data models does not solve the incomparability problem be-
cause the CPI in individual years before the year 2012 included data for different
components and time periods to measure perceived corruption across continents.
We believe that measuring corruption in the public sector by the CPI is suitable.
However, one cannot conclude from previous studies that corruption decreases
growth because the earlier version of the CPI is not comparable across time.

They go on to use two-stage least squares (2SLS) analysis with lagged corruption scores as
instrumental variables in a dynamic panel data model (Arellano and Bond, 1991; Blundell
and Bond, 1998) to estimate the effect of corruption on growth. Changes in TI CPI score
need to be valid measures of change in perceived corruption in order for their model to
generate valid causal inferences.
Some differences among country-level measures of corruption are probably attributable
to bias or measurement error. If individual measures are influenced by (a) systematic factors

2
that are not part of aggregate corruption or (b) random noise, then we must be able to filter
out these extraneous influences so that when we use these measures we are studying only
changes in corruption. If we don’t, any inferences that we draw using these indicators will be
contaminated by their biases and made more uncertain because of their noisy measurement.
There are also differences among corruption measures most likely rooted in conceptual
disagreement about what corruption is and how it should be measured. Where a scholar
has reason to strongly prefer one conceptualization over another, it is appropriate to use
the measure that embodies that concept instead of others. Yet all these measures claim to
be measures of the overall breadth and depth of corruption’s influence in a country, and
thus in principle should be tracking the same concept. Insomuch that there is a degree of
consensus about what constitutes corruption, measuring the level of corruption consistent
with that consensus enables us to conduct research whose conclusions are not as sensitive to
conceptual disagreement. Moreover, scholars’ use of any and all of these measures is typically
and explicitly justified on the basis of the degree to which each is comparable to others (for
examples, see Ko and Samajdar, 2010; Thomas, 2010; Langbein and Knack, 2010; Dincer
and Gunalp, 2012; Tabish and Jha, 2012; Mondo, 2016; McMann et al., 2016; Gründler and
Potrafke, 2019); thus comparing the measures is already something scholars regularly do in
order to validate their approach.
In this paper, we argue that the correlation among within-country changes of corruption
for different measures is an assessment of the degree to which those measures track the same
concept; this idea is sometimes called convergent validity. We use a data set containing
information about 199 countries from 1980-2020. The key variables in our analysis are five
influential measures of corruption: the TI CPI, WBGI CCE, BCI, ICRG, and the Varieties of
Democracy (V-Dem) political corruption index (Coppedge et al., 2021). Our analysis shows
that the five measures of corruption we study are all influenced by a shared underlying
corruption dimension, but frequently diverge in their assessment of within-country change

3
in corruption over time. We first explain the methodology behind these measures, as each
takes a different approach. Then, we describe a mathematical decomposition of time-series
cross-sectional (TSCS) measures that identify change over time within units. We use that
decomposition to construct panel-adjusted versions of each of the five measures of corruption
we study; our panel-adjusted measures are designed to accurately track changes in perceived
corruption within countries over time while ignoring systematic between-country differences
or worldwide annual shocks. We find that panel-adjusted versions of the CPI, WBGI, BCI,
V-Dem, and ICRG corruption measures are weakly correlated, but give very different answers
to simple and important questions about the trajectory of corruption within countries.
Finally, we present an alternative measurement model of corruption. Our model is similar
to the unobserved components model (UCM) behind the WBGI and BCI, but directly maps
onto principal component analysis and thereby allows us to extract a common component
of all five corruption perception variables corresponding to a latent concept that influences
them all. Most importantly, our approach imposes fewer assumptions on the structure of
measurement error than the UCM; it is robust to the presence of systematic bias among the
measures and error correlation across them. While all five measures load onto a common
factor in the expected way—that is, while all five measures seem to contain a signal corre-
sponding to the perception of aggregate corruption—this factor explains comparatively little
variance in the component scores. For all these reasons, we conclude that there are substan-
tively meaningful differences between these measures that are not ascribable to transient
noise and that will affect statistical inference.
Although the field’s interest in sub-national and sector-specific measures of corruption
is increasing (Heywood, 2014, p. 148), we agree with Mungiu-Pippidi and Fazekas (2020)
that valid and reliable country-level measures remain vitally important for the study of
corruption. We find their argument persuasive when they say (p. 24) that “the national
context—the level where the rules of the game are set—is crucial.” Exploiting sub-national

4
variation in corruption is advantageous in many ways, but a great deal of institutional
variation is between countries and over long time periods. If we think institutions affect
corruption, time-series cross-sectional data is a particularly appropriate place to study those
relationships. Furthermore, findings derived from within-country data have limited external
validity until demonstrated in a cross-national context. The factor score corruption measure
we create using principal component analysis, which extracts the common signal from all
five panel-adjusted measures of perceived corruption, allows for a more robust assessment of
within-country change in corruption compared to any of the individual component measures.
Our proposal does not answer every critique in the literature; for example, we do not address
the claim that perception-based measures do not predict less subjective proxies for corruption
(Donchev and Ujhelyi, 2014), a claim already disputed elsewhere (Charron, 2016). However,
we do believe that our measure provides a strong response to one very important criticism
of cross-national corruption measures, and we provide our factor score measurements as part
of the replication data so that researchers can use and improve upon them in future work.

Comparing Measures of Corruption

The five quantitative, country-level TSCS measures of corruption that we study are con-
structed using different methodologies and data sources. This fact has served as a basis for
criticism of their concept validity, particularly for measures that aggregate many disparate
assessments into a single index like the WBGI CCE (Thomas, 2010; Oman and Arndt, 2006,
p. 72). But it is well known that all these measures are highly correlated with one another,
with pairwise correlation coefficients over ρ = 0.9 (Table 3 in Standaert, 2015, p. 789; see
also Ko and Samajdar, 2010 pp. 520-521). This is hopeful news for their construct validity
and gives us reason to believe that, while different in some details, each measure captures
the overall pervasiveness of corruption in a country.
Unfortunately, correlation among these corruption measures “is almost completely driven

5
by their between-correlations (the correlation between the mean values for each country)”
(Standaert, 2015, p. 788) and not changes in corruption within countries over time. Criti-
cisms of the validity of within-country changes in corruption over time have focused on the
idea that “variations in reported levels of corruption are as likely to be a product of [...]
the methods that are used to create these measures, as they are to reflect actual levels of
corruption” (Heywood and Rose, 2014, p. 508). For this reason, even if these measures are
valid for cross-sectional studies, they may not be valid for changes in corruption within a
country over time.
The five corruption measures we study are created using very different procedures. To
highlight these differences, we describe each measure in depth in this section alongside some
of the criticisms that have been leveled at them and responses to those criticisms. We do not
comprehensively cover all criticisms that have been made of perception-based, country-year
corruption measures in our review; instead, we focus on criticisms that specifically pertain to
the validity of these measures to track changes in corruption over time.3 These descriptions
list the original scales of the variables, but we have rescaled them to range from 0-100 (with
larger numbers indicating more corruption) for all our analyses.

Transparency International’s Corruption Perceptions Index (TI CPI)4

The CPI is an extremely influential indicator of corruption widely used by scholars and
policymakers. According to Galtung (2006, p. 106), “The impact of the CPI has been
considerable. It has been credited as a factor that gave the issue of corruption ‘greater
international prominence’ (Florini, 1998).... The CPI has facilitated a qualitative shift in
the journalistic writing and public discourse on corruption.... This interest and awareness
3
The descriptions in this section are similar or identical to descriptions in the online appendix of Esarey and
Dalton (2023); these papers were written at the same time using (some of) the same variables.
4
Information about the CPI has been paraphrased from Transparency International (2016) and Transparency
International (2020)

6
of the CPI extends well beyond the business and financial press.” However, the use of
the CPI has become more controversial over time (Bello y Villarino, 2021; Heywood and
Rose, 2014). As its name explicitly says, the CPI measures perception of corruption and
not personal experience with corruption or convictions for corruption-related crimes. The
CPI is constructed by averaging at least three (but as many as thirteen) different corruption
scores taken from perception-based surveys and expert assessments of corruption in a given
country. The CPI targets perceived corruption in the public sector within a country and
compiles relevant data from multiple, independent sources.
The CPI has been a frequent target of criticism in part because its methodology did
not always enable accurate tracking of corruption trends within countries over time. In
response to that criticism, Transparency International altered its methodology in 2012 to
ensure that consistent and contemporaneous sources were used year after year (Transparency
International, 2012); for scores prior to 2012, the underlying sources of data changed from
year to year and information from several prior years was used to create the target year’s
score. Like all perception-based measures, TI CPI relies on judgments made by experts and
survey respondents; these sources may have conflicting understandings among themselves
about what corruption means or may have conceptualizations of what corruption means
that do not match the definition prevalent in the country itself. The CPI’s native scale
ranges from 0 (most corrupt) to 100 (least corrupt) and is available from 1995 onward.

World Bank Group’s Worldwide Governance Indicators (WBGI)5

The WBGI consists of six aggregated indicators of governance—control of corruption, voice

and accountability, rule of law, government effectiveness, political stability, and regulatory
quality—compiled from 30 data sources. The WBGI’s sources include many different surveys
(of the general population and of selected sub-groups) and expert assessments, including
5
Information about the WBGI has been paraphrased from Kaufmann, Kraay and Mastruzzi (2010).

7
some of the other measures of corruption we study here (such as the ICRG political risk of
corruption measure). It utilizes an Unobserved Components Model (UCM) to construct six
aggregated indicators of governance and estimate margins of error for each indicator. Of the
six indicators, our interest is in their measure of control of corruption, defined by Kaufmann,
Kraay and Mastruzzi (2010, p.4) as “the extent to which public power is exercised for private
gain, including both petty and grand forms of corruption, as well as ‘capture’ of the state
by elites and private interests.”
Some have argued that the six WBGI indicators are not empirically or conceptually dis-
tinct from each other (Langbein and Knack, 2010). In addition, the corruption measure
combines many disparate concepts of corruption into an amorphous and unidentifiable ag-
gregate (Apaza, 2009, p. 141), a criticism that may be applicable to many country-level
omnibus measures (such as the TI CPI). The WBGI’s creators reject these claims (Kauf-
mann, Kraay and Mastruzzi, 2007) in part based on the design of their UCM. The WBGI
has also been criticized because of the large standard errors associated with its governance
estimates; the uncertainty of these estimates makes it difficult to find statistically significant
differences in corruption scores within countries (Bello y Villarino, 2021). The WBGI’s orig-
inal scale ranges from -3 (least control over corruption, or highly corrupt) to 3 (most control
over corruption, or least corrupt), and is available for 1996, 1998, 2000, and 2002 onward.

Bayesian Corruption Index (BCI)6

The BCI is another index of the perception of overall corruption (abuse of public power for
private gain) within a country. It is built using an extension of the WBGI’s Unobserved
Components Model and based on 17 different surveys of countries’ inhabitants, business
executives, and government officials. Unlike the WBGI, the BCI implements a state-space
model that allows for both the overall level of corruption worldwide and any individual
6
Information about the BCI has been paraphrased from Standaert (2015).

8
country’s level of corruption to change over time; by contrast, the WBGI assumes that
global corruption levels remain constant over time. The underlying source data of the BCI
are entered without any ex-ante imputations, averaging, or other manipulations, thereby
avoiding selection biases introduced through the modeling choices of the index’s creator.
The BCI ranges between 0 (least corrupt) to 100 (most corrupt) in countries and is available
from 1984 onward (a larger time span than the CPI and WBGI).

Political Risk Service’s International Country Risk Guide (ICRG)7

The International Country Risk Guide provides political, economic, and financial risk rat-
ings to inform businesses about potential risks to their firms when operating within certain
countries. The corruption measure is a panel of experts’ assessment of the risk to businesses
and foreign investors that corruption presents in a given country; it may therefore be bi-
ased in favor of the attitudes, priorities, and viewpoints of businesses rather than individual
citizens or other organizations. Furthermore, the ICRG’s concept of corruption includes
concepts such as the prevalence of bribery, patronage, nepotism, extortion, and suspicious
ties between business and politics, some of which are excluded from the other measures;
the V-Dem measure of political corruption, for example, excludes patronage and clientelism
(Lindberg, Lo Bue and Sen, 2022). Because the ICRG may give different weights to the
various types of corruption going on in a state between years, it may be difficult to compare
the measure year-over-year even though it is designed for this purpose (Knack, 2007, p. 261).
As produced by the ICRG, the measure ranges between 0 (low political risk from corruption)
and 6 (high political risk from corruption) and is available for 1994 onward.
7
Information about the ICRG has been paraphrased from The PRS Group (2020)

9
Varieties of Democracies (V-Dem) Political Corruption Index8

The V-Dem project collects 470 measures related to democratic governance, each of which
is built using structured subjective assessments by country experts. Their index of political
corruption is based on averaging information from four subsidiary measures: (i) the public
sector corruption index, (ii) the executive corruption index, (iii) a measure of legislative
corruption, and (iv) a measure of judicial corruption. Each of these four sub-indexes is
in turn created from the output of an item response theory (IRT) model combining many
experts’ assessments about different aspects of corruption in the targeted sector. This IRT
model is designed to allow meaningful comparisons of country-level corruption (or any of the
four subsidiary measures) over time. The resulting composite measure of political corruption
ranges from 0 to 1, with 0 indicating low corruption, and is available in our data between
1980 and the present; the V-Dem data set goes back considerably further, with historical
assessments (made by scholars in the present day) going all the way back to 1789.

A Simple Example: Corruption in China and the United States

The inconsistency of these five measures of corruption can be illustrated with a simple
descriptive example. Consider Figure 1, which shows corruption measurements over time
for China (panel 1a) and the United States (panel 1b).9 For this graph, we subtracted the
value for each country-year from the country’s overall mean on that measure to emphasize
differences in the trajectory of within-country change over this time period. While the
measures clearly share commonalities, from year to year they often disagree on the direction
of change. For example, in China, the ICRG indicates a relatively stable corruption level
8
Information about the V-Dem has been paraphrased from Coppedge et al. (2021)
9
Our source for the five corruption measures we study are the WBGI dataset, the BCI dataset (via the
Quality of Government data set or QOG from Teorell et al., 2019), the V-Dem data set, and the ICRG
data set. Appendix A shows a full list of countries available in our data set. Some countries do not have
all corruption scores available for certain years and some indicators are only available for segments of the
time period; see Figure 6 in Appendix B for the availability of indicators over time.

10
between 2005 and 2018. But the other measures indicate a sharp decline in corruption. For
the United States, the V-Dem and BCI measures indicate stability in corruption levels over
the last forty years. By contrast, the WBGI shows dramatic growth in corruption between
the years 2000 and 2010.
The substantive upshot of this divergence is that it is difficult to answer even relatively
basic descriptive questions about influences on corruption in a country because different
measures do not agree. Consider one such question: was the Obama administration more or
less corrupt than other US presidential administrations between 1980 and 2020? To answer
that question, we can estimate a relatively simple time series model on corruption data from
the United States:

US corruption t = β0 + β1 × year t + β2 × Obama t + εt (1)

The variable Obama is a dummy for whether the observation t occurs between the years
of 2009 and 2016, inclusive.10 We need to consider the potential non-stationarity of this
time series to avoid the potential for spurious correlation; although corruption measures are
bounded and therefore by definition have finite mean and variance (unlike non-stationary
series), the augmented Dickey-Fuller and Phillips-Perron tests still fail to reject the null of
non-stationarity for this series (although the KPSS test fails to reject the null of stationarity
for the same series).11 As a robustness check, we also present results from a first-difference
model to eliminate any potential unit root.
Figure 2 presents the estimated coefficient and 95% confidence interval for the Obama
variable in the model of equation 1. The left panel uses dependent variables in levels, while
the right panel uses the first difference as the dependent variable. The particular corruption
10
As this is a single time series and we are estimating a trend term, there is no need to employ panel-adjusted
dependent variables for this model (described in the next section) as the results would not be different.
11
We conducted these tests using the aTSA library (Qiu, 2015).

11
 Figure 1: Two examples of divergent corruption trajectories

12
(a) China (b) United States

Figure 1a depicts the trajectory of six corruption perception measures (the five observed panel-adjusted vari-
ables) from 1980 to 2020 for China. Figure 1b depicts the trajectory of the same corruption measures from 1980
to 2020 for the United States of America. All measures shown are centered on the country mean. Variables
are scaled so that more positive numbers mean greater perceived corruption.
Figure 2: Estimated effect of the Obama administration (2009-2016) on corruption in the
United States

The left panel shows the coefficient and 95% confidence interval for the Obama
variable in the model of equation 1 using the dependent variable indicated on
the x-axis. Each measure theoretically ranges between 0-100 across all coun-
tries and time periods, with larger values indicating more corruption. The right
panel shows the same coefficient and 95% CI for the same model with the first-
differenced dependent variable.

13
perception measure used as the dependent variable is indicated on the x-axis of each graph.
Disturbingly, when studying the level of perceived corruption (the left panel of Figure 2),
the five measures give very different indications of the extent of corruption in the Obama
administration. If we interpret statistically insignificant results as being null findings (some-
what misleadingly; see Rainey, 2014), three of five measures report no difference between
the Obama administration and other US presidencies, one (the BCI) finds that the Obama
administration was more corrupt, and the last (V-Dem) finds that the Obama administra-
tion was less corrupt. None of the models using first-differenced corruption as a dependent
variable finds a statistically significant relationship between corruption and the Obama ad-
ministration, although the signs of the estimates still do not agree.

Panel Decomposition of Corruption Measures

Why do we see these discrepancies between corruption measures when studying within-
country changes over time? Although the correlation between the measures we study is
strong, this is attributable to differences in corruption between countries and not changes
in corruption within countries over time (Standaert, 2015). To show this, and to create a
measure specifically targeted at within country changes in corruption, we extract the average
differences in corruption between countries as well as any global trends or shocks in corruption
over time. We hereinafter refer to these measures as panel-adjusted.
Consider a theoretical decomposition of corruption measure Yit in a country i at time t
over a time period t ∈ {1...T }:
Yit = gi (t) + εit (2)

Each country i’s corruption level is given by a country-specific function gi (t) that represents
the trajectory of corruption over time t plus an added stochastic component εit ∼ f (µ, σ 2 )
with mean µ = 0 and variance σ 2 that represents random influences on corruption and/or
pure measurement noise. We can extract between-country differences in the corruption

14
measure by rewriting equation 2 as:

T T
!
1X 1X
Yit = gi (t) + gi (t) − gi (t) + εit (3)
T t=1 T t=1

= Ai + γi (t) + εit (4)

where Ai represents country i’s average corruption over the time period t ∈ {1...T } and
γi (t) is the de-meaned function representing its trajectory over time. We speculate that
the measurements of corruption we are studying (TI CPI, WBGI, BCI, V-Dem, and ICRG)
can distinguish countries’ average levels of corruption Ai from one another more easily than
countries’ changes in corruption over time.
Suppose there are any common global shifts in corruption, either due to measurement
noise or a genuine change in the overall level of corruption worldwide. In that case, we will
need to remove these impacts from the measure as well if we are trying to study country-
specific net changes in corruption.12 Removing that component is important if (as is most
common) we are studying influences on corruption that vary from country to country and are
not system-wide. It is also important to determine how much our corruption measures are
able to distinguish country-specific changes in corruption excluding overall global changes.
We therefore rewrite equation 4 as:

N N
!
1 X 1 X
Yit = Ai + γi (t) + γi (t) − γi (t) + εit (5)
N i=1 N i=1

= Ai + Pt + ψi (t) + εit (6)

12
This may or may not be necessary according to the objectives of the study and our assumptions about the
measurement process. If we believe that overall global trends in corruption measurement are indicative
of a time-varying bias in corruption ratings, then we would certainly want to remove that trend before
analysis (similar to an argument made about measures of democracy made by Little and Meng, 2023). On
the other hand, if we believe that there are average global changes in corruption and these changes are
relevant to the causal mechanism we were studying, we may wish not to extract global trends from the
data.

15
where Pt is the global average corruption measure for time t and ψi (t) represents the remain-
ing variance in country-specific corruption.
We are able to estimate the components of equation 6 with a fixed effects model:

N
X T
X
yit = α̂i Ii + π̂t Jt + ω̂it (7)
i=1 t=1

where yit is the observed value of a measure of corruption for country i in year t, α̂i is
the average value of the corruption measure across time in country i (and a measure of
Ai ), π̂t is the average corruption value across countries in year t (and a measure of Pt ),
I = {I1 , I2 , ..., IN } and J = {J1 , J2 , ..., JT } are vectors of dummy variables for countries and
years respectively, and all remaining variance in the corruption measure is in ω̂it . Thus ω̂it is
a measure of ψi (t) + εit , country-specific corruption plus (possibly systematic) error. These
estimates are consistent as long as ωit is uncorrelated with country and year (Wooldridge,
2010, pp. 300-301). Consequently, ω̂it is what we term a panel-adjusted measure of corruption
for country i at time t excluding between-country variation and worldwide trends.13
With this model, we can estimate ω̂it for the CPI, BCI, ICRG, WBGI, and V-Dem cor-
ruption measures using least-squares dummy variables regression.14 As before, all corruption
measures are set to a 0-100 scale with larger numbers indicating more corruption. We then
extract the estimated residuals from the model, ω̂it , to create the new panel-adjusted measure
of corruption with between-country differences and worldwide time trends removed.
Figure 3 reports correlations among the raw corruption measures and panel-adjusted
scores using the residuals from the fixed effects model in equation 7.15 As expected, the raw
annual measures are highly correlated with one another (as reported by Standaert, 2015),
13
αi and πt may be defined relative to a reference category (without loss of generality); for example, if an
overall intercept coefficient is estimated as a part of equation 7 then α1 and π1 may be fixed at zero as is
typical in panel dummy variable models.
14
All analyses are conducted using R 4.2.3 (R Core Team, 2023), in this case with the basic lm function.
15
For more detailed correlation figures between raw measures and fixed effect residual scores, see Table 2 in
Appendix C.

16
whereas the correlation between the panel-adjusted measures is weak (median ρ̂ = 0.238).
That is, the typical corruption measure only explains about 6% of the variation in any other
corruption measure once between-country differences and worldwide trends are removed.
The low correlation between panel-adjusted corruption measures indicates that they cannot
agree on how much a country’s corruption level changes over time.
The relatively weak correlation among panel-adjusted corruption measures, and their dis-
agreement in answering simple substantive questions about corruption, leads us to question
whether the CPI, BCI, ICRG, WBGI, and V-Dem measures all correspond to the same under-
standing or concept of “corruption.” We answer this question by proposing a relatively simple
measurement model for the latent concept of corruption that will, in turn, imply a method-
ology allowing us to decide whether these five measures map onto the same latent concept.
Although similar to the Unobserved Components Model laid out by Kaufmann, Kraay and
Mastruzzi (2010), our approach does not assume any structure for non-corruption-related
variance in the measures; in particular, error terms among different measures can be bi-
ased and/or correlated. We verify this methodology with a simulation study that generates
data from our theoretical model of corruption measurement and demonstrates that it can
accurately recover latent corruption from biased and noisy measures.

Measurement Model and Analysis

Suppose that we have K many observed panel-adjusted measures; in our case, K = 5. We
begin by assuming that each panel-adjusted measure k is a composite of latent corruption
(ψit from equation 6) and extraneous variance ξitk :

ωitk = δk ψit + ξitk (8)

This implies that, if we knew the true values of δk , ωitk , and ξit we could recover ψit . We do
have an estimate for ωitk , our panel-adjusted measure of corruption for each k, that we can

17
 Figure 3: The correlation among raw and panel-adjusted measures of corruption

18
(a) Raw Measures (b) Panel-Adjusted Measures

Figure 3a depicts the correlation between pairs of raw measures of corruption (named at the top and right
panels). Figure 3b depicts the correlation between panel-adjusted measures. The stars indicate statistical
significance: (∗ = p < 0.1, ∗∗ = p < 0.05, ∗ ∗ ∗ = p < 0.01).
plug in:

ω̂itk = δk ψit + ξitk − εitk (9)

ω̂itk = δk ψit + νitk (10)

where in the second line we set νitk = ξitk + εitk . Rewriting equation 10 in matrix form for
all observations in the data set, we get:

Ω̂ = ψδ T + ν (11)
N T ×K

where Ω̂ is the N T × K matrix of panel-adjusted corruption measures, ψ is an N T × 1 vector

of the latent true value of corruption, δ is a K × 1 vector mapping latent corruption into
observable measures, and ν is the N T × K matrix of remaining variance in Ω̂ not explained
by ψ.
As a model of the observed panel-adjusted corruption scores, equation 11 is simply a
restatement of principal components analysis (Bro and Smilde, 2014, pp. 2816-2817). PCA
chooses the vector δ T = {δ1 , δ2 , ..., δK } in equation 11 that maximizes the variance of ψ
subject to the constraint that K
P
k=1 δk = 1 (without loss of generality). Of course, there is

no a priori guarantee that the ψ̂ estimated by PCA is “really” latent corruption. But because
all our measures are designed to target the same concept (corruption), if all the observed
variables load strongly onto a single dimension in the expected way it is reasonable to infer
that this latent dimension is corruption and the degree to which ψ̂ explains all the observed
measures of corruption is a measure of the construct validity of those measures. Although we
have written equation 11 with only one common factor ψ, if the matrix Ω̂N T ×K is of full rank
then singular value decomposition of Ω̂N T ×K can produce K many orthogonal unobserved
component dimensions in descending order of the proportion of variance in ΩN T ×K that each

19
explains by making ψ into an N T × K matrix (with the K columns corresponding to K
principal components) and δ into a K × K matrix.
We perform probabilistic PCA analysis16 of both the raw (unadjusted) corruption mea-
sures17 as well as our panel-adjusted measures of corruption. If all measures do not load onto
a single dimension or that dimension does not explain most of the variance, the measures
may:

1. target different notions of corruption;

2. include concepts other than corruption, or;

3. might be highly contaminated by measurement error.

As we noted above, our measurement model is similar to the Unobserved Components

Model of Kaufmann, Kraay and Mastruzzi (2010). However, to enable estimating the com-
ponents of their model via averaging, Kaufmann, Kraay and Mastruzzi (2010) make the
additional assumptions that errors across sources are independent (E [νitj νitk ] = 0 for j 6= k)
 2 
and homoskedastic within a source (E νitj = σj2 ). PCA, by contrast, does not assume
any structure for ν other than that it is orthogonal (i.e., unrelated) to ψ. Thus we believe
our approach improves upon Kaufmann, Kraay and Mastruzzi (2010) by being more robust
to the probable dependence among measures across time and space pointed out by prior
criticisms of the WBGI CCE and other TSCS country-level corruption measures.
16
We use probabilistic PCA (PPCA) to account for the missingness of some corruption measures for some
country years. Specifically, we use the implementation of PPCA in the pcaMethods package (Stacklies
et al., 2007).
17
PPCA is performed on the raw measures after they have been all been rescaled to range between 0 and
100 with larger numbers indicating more corruption. Additionally, as is standard for PCA analysis, all
measures are further rescaled to set their means at zero and their standard deviations at one before PPCA
is performed.

20
Simulation Study

To verify that our PCA methodology can accurately recover within-country changes in latent
corruption over time using biased and noisy measures whose error terms are correlated,
we simulate the process with known parameters and verify that PCA can recover those
parameters. Each simulated country i ∈ {1, 2, ..., N } has a true latent corruption value ψit
at time t ∈ {1, 2, ..., T } given by:

ψit = Ai + τψ(i) t (12)

Ai ∼ U [−4, 4]

τψ(i) ∼ φ(µ = 0.3, σ = 0.3)

However, there are J-many latent features λj for j ∈ {1...J} that might contaminate
measures of corruption; this represents the possibility that the people and organizations
that construct measures may make biased assessments of corruption using irrelevant factors
(such as institutional history or cultural stereotypes) or could be using a conceptualization
of corruption that is different from the target (for example, petty or grand corruption only
instead of overall corruption). For the simulation, we set J = 2 and make:

λ(ij)t = Λij + τλ(ij) t (13)

Λij ∼ φ(µ = 0, σ = 0.1)

τλ(ij) ∼ φ(µ = 0, σ = 0.6)

These are “pure” bias components in that they are completely unrelated to the target cor-
ruption dimension ψ; components that capture a different conceptualization of corruption
would be correlated with but not identical to ψ. This construction is meant to mirror our

21
theoretical idea that there are large differences in corruption between countries and that (on
average) biases are zero, but within-country changes in corruption are much smaller than
trends in the bias components.
Measures of corruption are constructed by presuming that each is a weighted average of
both the target concept ψ and the bias dimensions {λ1 , λ2 }. So for country i at time t, the
kth measure M(it)k is:

2
X
M(it)k = pψ(k) (ψit ) + pj(k) (λ(ij)t ) + εit
j=1

where pψ(k) + p1(k) + p2(k) = 1. For the simulation, we create six measures, all of which set
pψ = 0.7. The first two measures set p1 = 0.3 and p2 = 0, the second two set p1 = 0 and
p2 = 0.3, and the final two set p1 = p2 = 0.15. Thus, we presume that corruption measures
are “good” in the sense that they largely capture the target component of corruption but
still contain substantial bias and noise. This also builds in the fact that extraneous variance
in the vector of measures at time t can be systematically related across measures owing to
shared biases among subsets of the measures.
Finally, we vary the covariance structure among the six simulated measures to assess the
robustness of our measurement strategy to correlated measurement error. Specifically, for
country i at time t, the six-element vector of error terms εT = {εit1 , εit2 , ..., εit6 } is distributed:

ε ∼ Φ (µ, Σ)
   
0 1 ρ ρ · · · ρ
   
0 ρ 1 ρ · · · ρ
µ =  .  , Σ = 0.5 ×  .
   
 ..   .. .. .. 
   . .
   
0 ρ ρ ··· ρ 1

22
We conduct separate simulations setting ρ ∈ {0, 0.2, 0.4, 0.6, 0.8, 0.95}.
We simulate this data generation process 10,000 times for N = 100 countries and T = 10
time periods, thereby constructing 10,000 simulated data sets. We then construct panel-
adjusted versions of all six measures in the data and create a PCA-based measure of latent
corruption using those six panel-adjusted measures. This enables us to determine whether
our procedure is accurate in a simulated situation where the true within-country changes in
corruption are known (unlike in real data, where they are not directly observable).
Figure 4 shows the results of our simulation. The figure presents 95% simulation intervals
for the correlation between within-country corruption and both raw measures (row 1), the
PCA factor score extracted from panel-adjusted measures (row 2), and the panel-adjusted
measures (row 3). Simulated within-country corruption is only weakly correlated with simu-
lated raw measures of corruption (before panel correction). Panel-adjusted corruption scores
and PCA-based scores in our simulation are both much more strongly correlated with within-
country corruption, with the exact degree of improvement for PCA-based scores dependent
on the level of correlation in error among the raw measures; the PCA-based measure is
typically correlated with within-country corruption at ≈ 0.88 when measurement errors are
uncorrelated (ρ = 0). Finally (and unlike the panel-adjusted measures) there is a strong
average correlation between raw measures of corruption and true corruption scores that in-
clude both between-country and within-country variance (row 4), just we found among the
actual (non-simulated) measures of corruption in Figure 3.
With these simulated results reinforcing the robustness of our methodology to common
criticisms of TSCS country-level corruption measures, we now turn to deploying this method-
ology on actual corruption measures (the CPI, BCI, ICRG, WBGI, and V-Dem measures)
taken from the aforementioned country-year data studying 199 countries from 1980-2020.

23
Figure 4: Simulating the recovery of within-country corruption scores using panel correc-
tion and PCA

True Within−Country Corruption vs.

Untransformed Measures

True Within−Country Corruption vs.

PCA Measure

True Within−Country Corruption vs.

Panel−Corrected Measures

True Corruption vs.

Untransformed Measures

0.25 0.50 0.75 1.00

correlation

error correlation 0 0.2 0.4 0.6 0.8 0.95

Results from 10,000 simulated data sets of measurement of latent within-country

corruption. Each data set simulates 100 countries over 10 years. Error bars indicate
95% simulation intervals. The color of the bar indicates the correlation among error
terms for the five simulated raw measures of corruption.

24
Results from Principal Components Analysis

Table 1 shows factor loadings for all corruption measures on the first two principal compo-
nents (PC1 and PC2) produced by PPCA on the five measures of corruption in our country-
year data; the row labeled R2 displays the proportion of variance in corruption scores that
is explained by each principal component. Among raw annual corruption scores (the first
two columns in Table 1), 90% of the variance in corruption measures is accounted for by a
single dimension (PC1); all corruption scores load positively on this dimension.18 Although
the dimensions extracted by PPCA do not have an intrinsic interpretation, the fact that all
corruption measures load positively on a single dimension suggests that they all map onto a
single concept: aggregate corruption.
For the panel-adjusted measures of corruption, PPCA identifies a PC1 dimension (shown
in the second two columns of Table 1) with factor loadings extremely similar to the PC1
dimension for raw corruption scores. However, PC1 only explains 39.5% of the variance in
the within-country change measures of corruption as compared to 90% for the raw measures.
The second dimension extracted by PPCA (PC2) explains much more variance for the panel-
adjusted measures compared to the second PPCA dimension for raw corruption scores. This
finding suggests that corruption measures do track a common component of change in cor-
ruption within a country over time, but do so less accurately than they track between-country
differences in corruption. Furthermore, there are other (substantively unidentified) common
factors shared by these measures that explain a considerable portion of their within-country
variance: PC2 explains over 21% of the variance in corruption for the panel-adjusted scores.
18
We multiplied some factor loading matrices by −1 to place all first principal component loadings in the
same direction.

25
 Table 1: Factor Loadings and R2 for Principal Components

raw panel-adjusted
PC1 PC2 PC1 PC2
VDEM 0.473 -0.362 0.580 -0.767
WBGI 0.463 -0.023 0.436 0.174
BCI 0.451 -0.369 0.473 0.570
ICRG 0.403 0.856 0.372 0.065
TI CPI 0.443 0.006 0.334 0.227
R2 0.900 0.050 0.395 0.218

The factor loadings on the first two principal components (PC1 and PC2) for each
corruption measure are listed in the rows. The R2 row displays the proportion of
variance in corruption measures that is explained by the principal component in the
column.

Conclusion
Based on our findings, the most widely used corruption perception measures all share a
common concept of aggregate country-level corruption. Yet this shared component is not
as influential on within-country changes in these measures as it is on the between-country
differences in their levels, and explains little variance of within-country changes in corruption.
We also identify other influences on these measures that are substantively difficult to classify
and do not influence all measures equally. As a result, corruption perception measures often
do not agree when measuring changes in corruption within a given country over time.
From these findings, we draw two conclusions. First, we believe that researchers must
be careful when interpreting results that crucially depend on within-country changes in cor-
ruption. Fortunately, caution about the use of country-level measures is already widespread
in the scholarly literature. Many researchers already check the robustness of their results by
using multiple measures of corruption to assess them. This is a reasonable practice. How-
ever, it is suboptimal in that each of the measures contains noise and bias components that
make it harder to draw valid inferences.

26
 Second, we believe that researchers can (a) extract a common factor from multiple corrup-
tion measures using probabilistic principal components analysis of panel-adjusted corruption
measures and (b) use the resulting scores as a measure of (change in) corruption within coun-
tries. Recall our analysis in Figure 2, where we assessed the relationship between country-
level corruption perception and Barack Obama’s presidency using the TI CPI, WBGI CCE,
BCI, ICRG, and V-Dem measures individually. In that original analysis, we found evidence
that the Obama administration was more, less, or equally corrupt (compared to other ad-
ministrations between 1980 and 2020) depending on the measure used. When we instead
use our factor score measures to repeat that analysis, we produce the result shown in Figure
5. As we would expect from our own experience living under the Obama administration,
both levels and differences of the PPCA-based corruption perception measure indicate that
corruption during the Obama presidency was statistically indistinguishable from corruption
during the other years in the sample.
Not only do the results of Figure 5 have stronger face validity than many of the results
in Figure 2, we believe this interpretation is bolstered by the methodological features of our
PPCA-based score relative to its individual measures. Specifically, our measurement model
is designed to extract common signals of change from year to year among these five measures,
ignoring global trends and time-constant differences between countries, while simultaneously
remaining robust to various forms of bias and error correlation that served as an important
basis of criticism in past work. We provide our PPCA-based corruption perception measures
for the international system between 1980 and 2020 as a part of the replication material for
this paper; our hope is to make our measure easy for future researchers to study, criticize,
and use in substantive work.

27
Figure 5: Estimated effect of the Obama administration (2009-2016) on corruption in the
United States using first principal component extracted from panel-adjusted measures of
corruption using PPCA

The panel shows the coefficient and 95% confidence interval for the Obama variable
in the model of equation 1 using the first principal component extracted using PPCA
from the panel-adjusted versions of CPI, BCI, ICRG, WBGI, and V-Dem corruption
measures. The x-axis shows whether levels or differences of this principal component
serve as the dependent variable. The level-dependent variable (PPCA scores) has
been rescaled so that it ranges between 0 and 100, and this rescaled variable is used
to define the first difference dependent variable.

28
References
Andersson, Staffan and Paul M Heywood. 2009. “The politics of perception: use and
abuse of Transparency International’s approach to measuring corruption.” Political Studies
57(4):746–767.

Apaza, Carmen R. 2009. “Measuring Governance and Corruption through the Worldwide
Governance Indicators: Critiques, Responses, and Ongoing Scholarly Discussion.” PS:
Political Science & Politics 42(1):139–143.

Arellano, Manuel and Stephen Bond. 1991. “Some tests of specification for panel data: Monte
Carlo evidence and an application to employment equations.” The Review of Economic
Studies 58(2):277–297.

Bello y Villarino, José-Miguel. 2021. “Measuring Corruption: A Critical Analysis of the

Existing Datasets and Their Suitability for Diachronic Transnational Research.” Social
Indicators Research 157(2):709–747.

Blundell, Richard and Stephen Bond. 1998. “Initial conditions and moment restrictions in
dynamic panel data models.” Journal of Econometrics 87(1):115–143.

Bro, Rasmus and Age K. Smilde. 2014. “Principal component analysis.” Analytical Methods
6:2812–2831.

Brooks, Graham, David Walsh, Chris Lewis and Hakkyong Kim. 2013. Measuring Corrup-
tion. In Preventing Corruption: Investigation, Enforcement, and Governance. Palgrave
Macmillan UK chapter 2.

Charron, Nicholas. 2016. “Do corruption measures have a perception problem? Assessing
the relationship between experiences and perceptions of corruption among citizens and
experts.” European Political Science Review 8(1):147–171.

Coppedge, Michael, John Gerring, Carl Henrik Knutsen et al. 2021. V-Dem Codebook
v11.1. Varieties of Democracy (V-Dem) Project. URL: https://www.v-dem.net/en/
data/data/v-dem-dataset-v111/.

Dincer, Oguzchan C. and Burak Gunalp. 2012. “Corruption and Income Inequality in the
United States.” Contemporary Economic Policy 30(2):283–292.

Donchev, Dilyan and Gergely Ujhelyi. 2014. “What Do Corruption Indices Measure?” Eco-
nomics & Politics 26(2):309–331.

Esarey, Justin and Maya Dalton. 2023. “The Changing Relationship between Gender and
Corruption.” Working Paper. URL: http://www.justinesarey.com/Corruption_and_
Gender_Over_Time__Dalton_and_Esarey.pdf.

Florini, Ann. 1998. “The End of Secrecy.” Foreign Policy 111:50–63.

29
Galtung, Fredrik. 2006. Measuring the Unmeasurable: Boundaries and Functions of (Macro)
Corruption Indices. In Measuring Corruption, ed. Charles J. G. Sampford, Arthur Shack-
lock, Carmel Connors and Fredrik Galtung. Burlington, VT: Ashgate Publishing, Ltd.
pp. 101–130.

Gründler, Klaus and Niklas Potrafke. 2019. “Corruption and economic growth: New empirical
evidence.” European Journal of Political Economy 60:101810. URL: https://dx.doi.org/
10.1016/j.ejpoleco.2019.08.001.

Hawken, Angela and Gerardo L. Munck. 2008. Measuring Corruption: A Critical Assessment
and a Proposal. In Tackling Corruption, Transforming Lives, ed. Anuradha K. Rajivan
and Ramesh Gampat. Macmillan India.

Heywood, Paul M. 2014. Measuring Corruption. In Routledge Handbook of Political Corrup-

tion. Routledge Handbooks chapter 10, pp. 137–153.

Heywood, Paul M. and Jonathan Rose. 2014. “‘Close but no cigar’: the measurement of
corruption.” Journal of Public Policy 34(3). Publisher: Cambridge University Press.

June, Raymond, Afroza Chowdhury, Nathaniel Heller and Jonathan Werve. 2015. “A User’s
Guide to Measuring Corruption.” United Nations Development Program. URL: https:
//www.undp.org/publications/users-guide-measuring-corruption.

Kaufmann, Daniel, Aart Kraay and Massimo Mastruzzi. 2007. “The Worldwide Gover-
nance Indicators Project: Answering the Critics.” World Bank Policy Research Working
Paper. URL: http://documents.worldbank.org/curated/en/979231468178138073/
The-worldwide-governance-indicators-project-answering-the-critics.

Kaufmann, Daniel, Aart Kraay and Massimo Mastruzzi. 2010. The Worldwide Governance
Indicators: Methodology and Analytical Issues. SSRN. URL: http://ssrn.com/paper=
1682130.

Knack, Stephen. 2007. “Measuring Corruption: A Critique of Indicators in Eastern Europe

and Central Asia.” Journal of Public Policy 27(03):255–291.

Ko, Kilkon and Ananya Samajdar. 2010. “Evaluation of international corruption indexes:
Should we believe them or not?” The Social Science Journal 47(3):508–540.

Langbein, Laura and Stephen Knack. 2010. “The Worldwide Governance Indicators: Six,
One, or None?” The Journal of Development Studies 46(2):350–370.

Lindberg, Staffan I., Maria C. Lo Bue and Kunal Sen. 2022. “Clientelism, corruption and
the rule of law.” World Development 158:105989.

Little, Andrew and Anne Meng. 2023. “Subjective and Objective Measurement of Democratic
Backsliding.” Social Science Research Network, January 18. Available at SSRN: https:
//ssrn.com/abstract=4327307 or http://dx.doi.org/10.2139/ssrn.4327307.

30
McMann, Kelly, Daniel Pemstein, Brigitte Seim, Jan Teorell and Staffan I Lindberg. 2016.
“Strategies of Validation: Assessing the Varieties of Democracy Corruption Data.” V-Dem
Working Paper Series 2016(23). URL: https://www.v-dem.net/media/publications/
v-dem_working_paper_2016_23.pdf accessed 3/19/2023.

Mondo, Bianca Vaz. 2016. “Measuring political corruption from audit results: a new panel of
Brazilian municipalities.” European Journal on Criminal Policy and Research 22:477–498.

Mungiu-Pippidi, Alina and Mihály Fazekas. 2020. How to Define and Measure Corruption.
In A research agenda for studies of corruption, ed. Alina Mungiu-Pippidi and Paul M
Heywood. Edward Elgar Publishing pp. 7–26.

Oman, Charles and Christiane Arndt. 2006. Uses and Abuses of Governance Indicators. De-
velopment Centre Studies OECD Publishing. URL: https://books.google.com/books?
id=dvoKJhaZvm8C.

Qiu, Debin. 2015. aTSA: Alternative Time Series Analysis. R package version 3.1.2.

R Core Team. 2023. R: A Language and Environment for Statistical Computing. Vienna,
Austria: R Foundation for Statistical Computing.
URL: https://www.R-project.org/

Rainey, Carlisle. 2014. “Arguing for a Negligible Effect.” American Journal of Political
Science 58(4):1083–1091.

Sampford, Charles, Arthur Shacklock, Carmel Connors and Fredrik Galtung, eds. 2006.
Measuring Corruption. Ashgate Publishing Company.

Stacklies, Wolfram, Henning Redestig, Matthias Scholz, Dirk Walther and Joachim Selbig.
2007. “pcaMethods – a Bioconductor package providing PCA methods for incomplete
data.” Bioinformatics 23:1164–1167.

Standaert, Samuel. 2015. “Divining the level of corruption: A Bayesian state-space approach.”
Journal of Comparative Economics 43(3):782–803. URL: https://doi.org/10.1016/j.
jce.2014.05.007.

Tabish, S.Z.S. and Kumar Neeraj Jha. 2012. “The impact of anti-corruption strategies on
corruption free performance in public construction projects.” Construction Management
& Economics 30(1):21–35.

Teorell, Jan, Stefan Dahlberg, Sören Holmberg, Bo Rothstein, Natalia Alvarado Pachon
and Richard Svensson. 2019. “The Quality of Government Dataset, version Jan19.” URL:
http://www.qog.pol.gu.se.

The PRS Group. 2020. “International Country Risk Guide: Corruption

Risk Measure.” URL: https://www.prsgroup.com/explore-our-products/
international-country-risk-guide/.

31
Thomas, M. A. 2010. “What Do the Worldwide Governance Indicators Measure?” European
Journal of Development Research 22(1):31–54.

Transparency International. 2012. “Corruption Perceptions Index 2012: Technical Method-

ology Note.” URL: https://tinyurl.com/56a4avwp accessed 7/31/2023.

Transparency International. 2016. “Technical Methodology Note.” URL: https://bit.ly/

2Iu9IrU accessed 6/20/2019.

Transparency International. 2020. “Corruption Perceptions Index Overview.” URL: https:

//www.transparency.org/en/cpi/2020/index/nzl.

Treisman, Daniel. 2007. “What Have We Learned About the Causes of Corruption
from Ten Years of Cross-National Empirical Research?” Annual Review of Politi-
cal Science 10(1):211–244. URL: http://www.annualreviews.org/doi/abs/10.1146/
annurev.polisci.10.081205.095418.

Warren, Mark E. 2004. “What Does Corruption Mean in a Democracy?” American Journal
of Political Science 48(2):328–343.

Wooldridge, Jeffrey M. 2010. Econometric Analysis of Cross Section and Panel Data. Second
ed. MIT Press.

32
Appendices
A List of Countries with Available Data from 1980-2020
Afghanistan, Albania, Algeria, Andorra, Angola, Antigua and Barbuda, Azerbaijan, Ar-
gentina, Australia, Austria, Bahamas, Bahrain, Bangladesh, Armenia, Barbados, Belgium,
Bhutan, Bolivia, Bosnia and Herzegovina, Botswana, Brazil, Belize, Solomon Islands, Brunei,
Bulgaria, Myanmar, Burundi, Belarus, Cambodia, Cameroon, Canada, Cape Verde, Cen-
tral African Republic, Sri Lanka, Chad, Chile, China, Taiwan, Colombia, Comoros, Congo,
Congo, Democratic Republic, Costa Rica, Croatia, Cuba, Cyprus (1975-), Czechoslovakia,
Czech Republic, Benin, Denmark, Dominica, Dominican Republic, Ecuador, El Salvador,
Equatorial Guinea, Ethiopia (-1992), Ethiopia (1993-), Eritrea, Estonia, Fiji, Finland, France
(1963-), Djibouti, Gabon, Georgia, Gambia, Germany, Germany, East, Germany, West,
Ghana, Kiribati, Greece, Grenada, Guatemala, Guinea, Guyana, Haiti, Honduras, Hun-
gary, Iceland, India, Indonesia, Iran, Iraq, Ireland, Israel, Italy, Cote d’Ivoire, Jamaica,
Japan, Kazakhstan, Jordan, Kenya, Korea, North, Korea, South, Kuwait, Kyrgyzstan, Laos,
Lebanon, Lesotho, Latvia, Liberia, Libya, Liechtenstein, Lithuania, Luxembourg, Mada-
gascar, Malawi, Malaysia (1966-), Maldives, Mali, Malta, Mauritania, Mauritius, Mexico,
Mongolia, Moldova, Montenegro, Morocco, Mozambique, Oman, Namibia, Nauru, Nepal,
Netherlands, Vanuatu, New Zealand, Nicaragua, Niger, Nigeria, Norway, Micronesia, Mar-
shall Islands, Palau, Pakistan (1971-), Panama, Papua New Guinea, Paraguay, Peru, Philip-
pines, Poland, Portugal, Guinea-Bissau, Timor-Leste, Qatar, Romania, Russia, Rwanda, St
Kitts and Nevis, St Lucia, St Vincent and the Grenadines, Sao Tome and Principe, Saudi
Arabia, Senegal, Serbia, Seychelles, Sierra Leone, Singapore, Slovakia, Vietnam, Slovenia,
Somalia, South Africa, Zimbabwe, Spain, South Sudan, Sudan (2012-), Sudan (-2011), Suri-
name, Eswatini (former Swaziland), Sweden, Switzerland, Syria, Tajikistan, Thailand, Togo,
Tonga, Trinidad and Tobago, United Arab Emirates, Tunisia, Turkey, Turkmenistan, Tu-
valu, Uganda, Ukraine, North Macedonia, USSR, Egypt, United Kingdom, Tanzania, United
States, Burkina Faso, Uruguay, Uzbekistan, Venezuela, Samoa, Yemen, Serbia and Montene-
gro, Zambia

33
B Corruption Measure Availability by Year
Figure 6: Availability of Corruption Measures by Year

The availability of each of the five corruption measures during between the years 1980
to 2020, inclusive.

34
C Annual Correlation Table
Table 2: Correlation Among Corruption Measures

Measure Data Type VDEM WBGI BCI ICRG TI CPI

VDEM raw score, annual 1.000 0.901 0.804 0.709 0.869
panel-adjusted, annual 1.000 0.272 0.109 0.189 0.121
panel-adjusted, decade avg. 1.000 0.381 0.144 0.272 0.202
WBGI raw score, annual 0.901 1.000 0.907 0.872 0.974
panel-adjusted, annual 0.272 1.000 0.325 0.308 0.540
panel-adjusted, decade avg. 0.381 1.000 0.392 0.417 0.573
BCI raw score, annual 0.804 0.907 1.000 0.700 0.919
panel-adjusted, annual 0.109 0.325 1.000 0.170 0.251
panel-adjusted, decade avg. 0.144 0.392 1.000 0.228 0.246
ICRG raw score, annual 0.709 0.872 0.700 1.000 0.873
panel-adjusted, annual 0.189 0.308 0.170 1.000 0.225
panel-adjusted, decade avg. 0.272 0.417 0.228 1.000 0.259
TI CPI raw score, annual 0.869 0.974 0.919 0.873 1.000
panel-adjusted, annual 0.121 0.540 0.251 0.225 1.000
panel-adjusted, decade avg. 0.202 0.573 0.246 0.259 1.000
# of country-years 6802 3977 5639 4662 3650

Correlation among corruption measures for raw annual scores, panel-adjusted annual
scores, and panel-adjusted decennial average scores.

35