What corruption means informs what and how anticorruption reformers reform.
Unfortunately, Transparency International’s Corruption Perceptions Index (CPI) dodges this important issue by averaging together the responses from polls employing competing definitions of “corruption.” This is a problem because different types of corruption have different causes, have different effects, and require different types of remedies. Transparency International should disaggregate its index of perceived “corruption” into two distinct indices: one for perceptions of illegal corruption, and one for primarily legal (but distrust-generating) conduct, which could fairly be characterized as institutional corruption. This change would make the CPI more precise, better educating the press, public, and policymakers who rely on it.
According to Transparency International, the CPI “ranks countries and territories based on how corrupt their public sector is perceived to be.” While the reference to perceptions implicates Matthew’s “public opinion” definition of corruption, the CPI is constructed by averaging findings from thirteen polls employing disparate definitions of corruption, or leaving the term undefined. These polls ask different panels of experts (which might be problematic in its own right) a wide variety of questions. These questions all employ the term “corruption,” but its meaning varies quite considerably among them. By my count, these questions use four different definitions of corruption:
- A public-interest definition: asking “how common is diversion of public funds to companies, individuals or groups due to corruption?” and how often do officials use “public office for private gain?”
- A legal definition: inquiring about the extent to which illegal corruption is prosecuted and the prevalence of “bribing and corruption.”
- A public opinion definition (although in this case, the “public” are exclusively experts): having experts assess the extent of perceived “state capture by narrow vested interests,” “suspiciously close ties between politics and business,” and the excessiveness of “bureaucratic regulations,” all of which presumably decrease the public’s trust in government.
- Undefined: asking about the extent of “corruption” in the public sector, which might implicate any of the above three definitions.
In light of evidence that the answers to the above questions are highly correlated (demonstrated by a European Commission study), one might argue that the inconsistency of the definitions is irrelevant. However, the answers to the questions are far from perfectly correlated; in fact, as the EC study noted, “if the countries’ classifications in the thirteen sources were to be taken at face value . . . no country is classified as better off than another country on all common sources.” The EC study applauds this finding, stating that it shows that the CPI “reconciles” the differing definitions of corruption, but in fact the CPI does not “reconcile” these differences at all–it merely averages the scores, effectively splitting the difference between competing definitions. This is problematic because the methods for alleviating the different corruptions vary: preventing bribery-type behavior and “state capture by narrow vested interests” require different types of legislation. It is the difference between an anti-bribery law and a campaign finance bill.
Of course, when constructing an index one must always sacrifice some degree of nuance for simplicity. But the CPI could be substantially improved by distinguishing between perceptions of corruption that implicate (1) illegal, or close to illegal, conduct, and (2) legal, or mostly legal, conduct. Transparency International should classify the polls it already uses as falling into one category or the other, and ditch the polls that do not define corruption. This would preserve a relatively simple index, but still allow observers, journalists, and analysts to appreciate the distinction between very different types of “corruption”, which may have quite different manifestations and call for quite different policy responses.
I feel like I should know the answer to this question, but I don’t, so I’ll throw it out there:
Has anyone performed a principle components analysis on the 13 separate surveys that TI uses to construct the CPI? If Michael is correct and there are really two main concepts of corruption here, then the principal components analysis should show that these 13 variables load onto two main factors, which (when we look more closely at the data) should correspond to the two types of “corruption” Michael describes.
On the other hand, if one factor seems to explain most of the variance, that would be evidence against Michael’s argument in this post. (Likewise, if we need more that two factors to explain most of the variance, or if the factors we identify don’t seem to match substantively with the two corruption concepts Michael identifies, that would also be evidence against his argument.)
This seems like an eminently do-able exercise (so long has one has the data and the right statistical training and software). Has anybody done it?
The statistical analysis would be useful, but it would still leave me with two issues. First, in addition to using two different definitions of corruption, multiple surveys TI uses employ “corruption” without defining it at all, so my thesis is not that there would be two load factors (my thesis IS that, if just the explicit definitions of each type were used, then they would load onto two factors). More generally, if the two variables (my two “corruption”s) are highly correlated, they will show up as one load factor (i.e., one factor WILL explain most of the variance in the data), but I don’t think that would defeat my analysis: even if the two corruptions are highly correlated, so long as (1) the correlation is less than one and (2) each corruption requires a radically different policy response, then there will be benefits to separating them. Further, currently correlated-in-a-cross-section-of-data variables could be less (or not) correlated in a single country (over time-series) in response to policy changes.
Dodging Matthew’s question (because I have no answer), I’d like to thank you for a powerful and persuasive post, Michael. I think disaggregating institutional and illegal corruption is helpful, but I’ve got a very basic question. Who is TI’s target audience? If the goal is generating specific policies, disaggregation is helpful; if the goal is to move people to think generally about corruption, disaggregation can run counter to that end.
Michael, that’s a great point. When I read your post, I immediately thought, “But if they’re using the same polls for each country, can’t this be seen as a valid aggregation under some kind of lowest-common-denominator theory of corruption?” In other words, couldn’t this all be boiled down to perception of corruption, or something like that? At that point, I don’t think the index would be that helpful, but at least it could be used for relative comparisons.
But, taking a quick look at the data, it seems that they don’t even reliably use the same polls across countries. Sometimes they aggregate as few as 3 and as many as 10 of the thirteen polls. Given the inconsistency in definitions of corruption AND the inconsistency in which polls they use, it seems important to take the results with a big grain of salt.
But, then, this opens up another interesting question — which is why does the heat map look kind of like what I expected (albeit with some outliers)? Perhaps this just implies that different kinds of corruption are correlated enough that the data is largely accurate, despite the methodological failings.
In fairness to TI, the researchers who have worked on creating the CPI — particularly Johann Graf Lambsdorff — have discussed some of these methodological issues, though I don’t think they’ve really engaged with Michael’s main conceptual point. it may be worth checking out Lambsdorff’s 2001 “How Precise Are Perceived Levels of Corruption?” paper, which was published as a background paper with that year’s CPI, and his 2006 chapter, “Measuring Corruption – The Validity and Precision of Subjective Indicators (CPI”)”, which appeared in the book Measuring Corruption.
Matthew, thanks for pointing me to those sources! You’re absolutely right that the CPI authors have discussed some of these issues — for example, the 2001 paper points out that the number of sources (which I mentioned in my last comment) affects the confidence level of the CPI score, which makes sense and is reflected in their 2013 chart. My comment was more an attempt to draw out a minor corollary to Michael’s point. I apologize if I was unclear; I’ll try to clarify it a bit.
Essentially, I was trying to say that the variable combination of the polls means that, apart from the question of whether the polls use consistent definitions, the poll aggregation itself is not necessarily apples-to-apples, either, precisely because there might be different categories of polls. For example, if they use thirteen polls, let’s say that six use definitions of corruption more aligned with venal corruption (bribery, etc.) and seven use institutional corruption definitions. If a country has been analyzed by only 3 or 4 polls (about 35 countries on the list), it’s entirely possible that all of those polls fit into only one of the two categories of polls, or are very heavily weighted toward one of the two categories. Using very different combinations of polls complicates the problem Michael pointed out even further. I know this is only a very modest point, but I wanted to make sure it’s clear. Thanks again for pointing me to those useful resources.
Yes, you’re right. Your point – and Michael’s point in the original post – is that aggregating (whether by averaging or by some fancier statistical technique) presumes that the variables we’re aggregating are all measuring more or less the same thing, but with some idiosyncratic measurement error. But if the measures are really measuring very different things – and if different measures are available for different units – we can run into trouble.