Corruption around the world: an analysis by partial least squares—structural equation modeling

Searching for the word “corruption” in the titles of the documents indexed by Scopus database (Subject Areas: Social Sciences; Economics, Econometrics and Finance), returns 366 responses between 1989 and 1998, 1480 in the 1999–2008 period and 5005 in the decade running from 2009 to 2018.

It is beyond the scope of the present research to examine the relationship between lobbying and illegal corruption. For a review of that literature, see Lambsdorff (2002), Damania et al. (2004), Campos and Giovannoni (2007), Gokcekus and Sonan (2017) and Goldberg (2018).

The deduction is based on analysis of Corruption Perceptions Index of Transparency International and the Control of Corruption variable of World Bank in the Sweden context. Specifically, “In such settings, bribery is more likely only the tip of the corruption iceberg, and undue influence and conflicts of interest are more frequent occurrences” (Andersson 2017, p. 70).

Kaufmann and Vicente (2011) point out that, mainly for developed countries, inadequate empirical attention has been paid to legal types of corruption. Recently, Gokcekus and Sonan (2017) and Dincer and Johnston (2019) contribute to filling that gap by estimating the sizes of and the relationship between legal and illegal corruption in a cross-state panel for the United States. Unfortunately, at this time, no cross-country panel data are available for extending their analysis to a global level.

An active debate is underway about whether legal (e.g., lobbying, political contribution) and illegal corruption are substitutes or complements; the results are still inconclusive (Shepsle 2017; Goldberg 2018). The main argument that they are substitutes relies on the idea that lobbying enables the lobbyist to change the rules, thus making corruption redundant (Harstad and Svensson 2011). The rationale that they are complements relies on the idea that legal and illegal corruption may be considered to be two sides of the same coin: on the one hand, legal corruption may be seen as a long-term investment aimed at influencing politicians to change the rules of the game; on the other hand, illegal corruption may be considered to be a short-term investment, directed to influencing public officials to find ways around the existing rules (Gokcekus and Sonan 2017).

Similar to the research herein, Dreher et al. (2007) estimate an index of perceived corruption with structural equation modeling—however, several differences arise in terms of: (1) estimation method—they estimate the model by a covariance-based approach, while I apply a PLS approach; (2) model specification—they estimate corruption with a multiple indicators and multiple causes (MIMIC) model, while I apply a broader structural model specification; (3) exhaustiveness of measurement and structural models - they define one latent variable (i.e., corruption) with five observable causes and four observable indicators, while I define 11 latent variables and, for each of those constructs, I specify a distinct measurement model, implying 47 manifest variables; (4) extensiveness of corruption indexes—Dreher et al. (2007) estimate an index of corruption that covers 100 countries over the 1976–1997 period; the index herein covers 165 countries over the 1995–2016 period.

In chronological order, I refer the reader to Ades and Di Tella (1997), Tanzi (1998), Rose-Ackerman (1999), Jain (2001), Aidt (2003), Svensson (2005), Lambsdorff (2006, 2007), Serra (2006), Treisman (2007, 2015), Goel and Nelson (2010), Enste and Heldman (2017) and Dimant and Tosato (2018).

More specifically, the model includes 20 latent constructs but 11 of these constructs have a single indicator with a loading coefficient fixed equal to 1 (see Table 2). Accordingly, those 11 (formative) constructs are equal to their corresponding single manifest indicators. These specifications of measurement models make it possible to estimate the path coefficients of observable variables (i.e., oil rents, decentralization, colonial and religion dummies—see Table 3) of structural models that, by definition, only include latent constructs.

Extensive reviews of the PLS approach to SEM are given in Chin (1998), Tenenhaus and Esposito Vinzi (2005), Esposito Vinzi et al. (2010a), Hair et al. (2016, 2017, 2019) and Faizan et al. (2018). The benefits and limitations of partial least squares path modeling (PLS) is still an open issue. On opposite side of the debate is Rönkkö et al. (2016).

Specifically (1) I aim to predict an index of perceived corruption; (2) the network of relationships between corruption and its potential economic, cultural, and institutional determinants is complex; (3) the specified model includes more formatively measured constructs; (4) the availability of several alternative indicators for measuring variables that intrinsically are unobservable and/or multidimensional and (5) the violation of the multivariate normality assumption.

That option deletes those observations for which values are missing in each pair of manifest variables.

The estimates are calculated by the “SmartPLS 3.0” software developed by Ringle et al. (2015).

A controversial issue of bias (and potential remedies) arises when use PLS–SEM to estimate reflective models. According to Sarstedt et al. (2016), the PLS algorithm is preferable to CB and PLSc, when it is not known whether the data's nature is common factor- or composite-based. Other studies (e.g., Dijkstra and Henseler 2015), state that PLSc is preferable to PLS. In the following, I choose to report PLS estimates instead of consistent PLS estimates (PLSc) and a bootstrapping routine applied to correct estimated coefficients on the reflective constructs (Dijkstra and Henseler 2015). That choice is supported by evidence that the differences between findings based on PLSc and PLS estimates are negligible and that the latent scores used to calculate the S-CPI are not affected by the decision.

To save space, I omit reporting the matrices. In brief, the analysis reveals that the hypothesis of discriminant validity holds for estimated model 1 (2) because only 2 (1) HTMT values of 153 (28) estimated ratios exceed the threshold.

R²s larger than 0.75, 0.50 and 0.25 indicate large, medium and moderate amounts of explained variations in endogenous construct.

A Q² exceeding 0.5 reveals the large predictive relevance of given latent variables, while when Q² is negative, no evidence of predictive relevance is found (Cohen 1988).

The rule of thumb is that if PLS yields a larger RMSE than LM for all, the majority of, the minority of, or the same number, or none of the observed indicators, then PLS has no, low, medium, or high predictive power, respectively.

The model with the smallest BICs and AICs is preferred.

However, because of the robustness of results across similar specifications, S-CPI scores are not significantly affected by that choice.

Although the different treatments of missing values don’t markedly change the rankings of countries between the S–CPI indexes estimated by the original dataset (MV) and the datasets with replacement (I and IFB) and their correlations are quite high (99.7% and 98.9%), the S–CPI scores are biased because missing values are more prevalent during the first decade of time range (1995–2005) and during the last 2 years of the sample (2015–2016).

The annual estimates (1995–2016 for 165 countries) of S-CPI are reported in Appendix A.2.

The original indexes are standardized in order to range over the same scale of the S–CPI (i.e., 0–100).

In order to fulfil the requirements for conducting the IPMA, I have taken the total effects in absolute values, such that higher values represent positive effects for the meaning of the key latent construct. In Fig. 2 the original signs of the total effect on perceived corruption are reported in parentheses.