Study Design: Sample, Data, Measures, and Empirical Strategies

Estimating the effects of politicians’ salaries is difficult because of endogeneity problems arising from a variety of strategic and contextual factors that determine salary levels. To address these challenges, we searched for all instances in countries in the European Union of changes in salaries of local government executives (led by mayors and co-governed by municipal councils)Footnote ¹ tied to the population of their municipalities. This set-up is advantageous because the formula-based determination of remuneration provides a plausibly exogenous source of variation in salaries. We identified ninety-seven such instances where salaries change discontinuously at discrete population thresholds in eleven countries: Austria, Belgium, France, Hungary, Italy, Lithuania, Portugal, Romania, Slovakia, Slovenia, and Spain.Footnote ² All thresholds are listed in Supplementary Appendix (SA) Section A.

This approach has been utilized in prior research studying the effect of salaries on governance outcomes (Klašnja Reference Klašnja2015; Pulejo and Querub (Reference Pulejo and Pablo2026); Pique Reference Pique2019).Footnote ³ Our contribution is to draw on a significantly larger number of cases, as our sample comprises close to 18,000 municipalities and covers countries accounting for more than half of the European Union’s population and half of its gross domestic product (GDP). The breadth of our sample gives us rich and meaningful variation in salaries. Panel (a) of Figure 1 shows that the average increase in mayoral monthly salary across all salary thresholds is approximately 15 per cent (from about Purchasing Power Parity [PPP] €2,650 to €3,050). Salary jumps vary between less than 1 per cent and 95 per cent, with an interquartile range of 5–30 per cent (SA Figure C2).

Figure 1. RD plots of change in average mayoral salary and corruption risk index at the cut-off. Notes: panel (a) includes all population-based salary thresholds; panel (b) includes only the unique thresholds described in the text. Dots are averages of the y variable within small ranges (or ‘bins’) of the x (or ‘running’) variable, which represents the per cent distance of a town to the closest salary threshold (indicated by the vertical dashed line); lines on each side of the discontinuity are cubic best-fit curves fit to the data (separately on the left and right of the discontinuity). Bins of the running variable are determined with the data-driven procedure by Calonico et al. (Reference Calonico, Matias and Titiunik2015).

Our focus on local executives is warranted on more than just research design grounds. Local governments in the European Union have substantial policy responsibilities in the provision of public goods and services, such as the maintenance of local infrastructure and utilities and the administration of local safety, health, education, and social protection (Kuhlmann et al. Reference Kuhlmann, Dumas and Heuberger2022). In the countries in our sample, local governments account for a fifth of total public expenditures and revenues, two-fifths of public investments, and a third of the total value of government contracting.Footnote ⁴ Local executives are crucial in formulating and implementing these policies, including public procurement, which is highly discretionary and thus subject to considerable corruption risks. SA Section B gives an overview of the organization, resources, and responsibilities of local executives in each country, highlighting their central policy roles and powers.

Given the local politicians’ roles in the allocation of government resources, we focus on measuring corruption in local government spending. We assembled a large original dataset of more than 2.4 million public procurement contracts in the countries in our sample over the period 2006–20. About a third of government spending in the European Union is implemented through public procurement – around €2 trillion annually.Footnote ⁵ We collected the data directly from official government sources and open data repositories using web-scraping algorithms (Fazekas et al. Reference Fazekas, Tóth, Abdou and Al-Shaibani2024).Footnote ⁶ Using the information on contracting authorities, we focus on tenders administered by local governments (accounting for around a third of all procurement), geocoding each contract and linking it to the data on local politicians’ salaries and salary thresholds. SA Table C1 gives details for each country sample.

We define six corruption risk indicators that approximate separate, but often correlated, aspects of corrupt deals in public procurement. In terms of procurement procedure, we track: (a) the use of non-open (that is, non-competitive) tender procedures, (b) the absence of call-for-tender announcements, (c) excessively short advertisement periods, and (d) excessively short award decision periods. With respect to tender outcomes, we track two indicators: presence of only a single bidder, and the spending concentration of contracting authorities (that is, local market concentration). In isolation, these indicators may reflect idiosyncratic, benign tender characteristics rather than corruption. To reduce the potential for such misinterpretation, we average the indicators into a composite procurement Corruption Risk Index (CRI) that varies between 0 and 1, with higher values indicating higher risk. Therefore, we assume that a contract characterized by multiple red flags is more likely to represent a corrupt tender. The index varies considerably between and within countries in our sample. SA Section C.1.2 provides more details for each indicator and the composite index.

The index is comparable across countries and has been empirically validated. For example, it correlates well with survey-based measures of corruption and is strongly predictive of overpricing in procurement tenders (Fazekas et al. Reference Fazekas, Tóth and King2016; Fazekas and Kocsis Reference Fazekas and Kocsis2020). The CRI therefore hues closer to measuring resource misallocation than the measures in much of the prior literature on salaries, in which some of the outcomes are only indirectly related to or themselves influenced by corruption, such as government size, local investment, or political competition (Ferraz and Finan Reference Ferraz and Finan2009; Gagliarducci and Nannicini Reference Gagliarducci and Nannicini2013; Pique Reference Pique2019). Others also use procurement-based performance or manipulation indicators (Klašnja Reference Klašnja2015; Pulejo and Querub ın Reference Pulejo and Pablo2026); however, unlike the CRI, some are context-specific and cannot be constructed more broadly (for example, bunching around regulation-specific contract value thresholds).

Using this data, our first empirical strategy is to estimate the following fixed effects model:

(1) $${Y_{imtc}} = {\rm{ }}{\alpha _{imtc}} + {\rm{ }}\beta {S_{mtc}} + {\rm{ }}\delta {X_{mtc}} + {\rm{ }}{\tau _{mtc}} + {\rm{ }}{\xi _m} + {\rm{ }}{\eta _t} + {\epsilon _{imtc}}$$

where Y is the CRI for contract i in municipality m in year t and country c; S is the mayoral salary in the same municipality and year (expressed in base-2020 PPP euros and logged), X is the municipal population (logged), τ indicates the closest population-based salary threshold, and ξ and η are the municipality and year fixed effects, respectively. We estimate the model with ordinary least squares (OLS) and cluster the standard errors at the municipality level.

The key coefficient, $\hat \beta$ , utilizes both the within- and across-municipality (and across-threshold) variation in mayoral salaries,Footnote ⁷ allowing us to use our large dataset to comprehensively describe the association between the population-driven variation in salary and the CRI. However, there may be various unobserved, time-varying characteristics of municipalities (or their politicians) that the model may not account for.Footnote ⁸ To overcome that challenge, we exploit the discontinuous nature of the increases in salary to apply a regression discontinuity (RD) design. This approach aims to estimate the causal effect of salary by analyzing corruption outcomes in plausibly comparable municipalities very close to the salary thresholds. Our RD model is as follows:

(2) $${Y_{imtc}} = {\rm{ }}{\alpha _{imtc}} + {\rm{ }}\theta 1[{{\tilde X} } \ge 0]_{mtc} + {\rm{ }}\gamma f({\tilde X)_{mtc}} + {\rm{ }}\lambda \ 1[\tilde X{\rm{ }} \ge 0]]_{mtc}{\rm{ }} \times f{(\tilde X)_{mtc}} + {\rm{ }}{\psi _c} + {\rm{ }}{\eta _t} + {\rm{ }}{\epsilon _{imtc}}$$

where variables with the same notation as in Equation 1 have the same interpretation; $\tilde{X} = {X - x_0 \over x_0}$ is a municipality’s percent distance in population X to the closest threshold x ₀;Footnote ⁹ and ψ are the country dummies. $\hat \theta$ captures the effect of a salary increase at the appropriate population-based threshold. Since we have thresholds at various population sizes, we normalize the distance of each municipality to a threshold so that the cut-off is zero for all municipalities. Therefore, $\hat \theta$ represents a weighted average of the local average treatment effects across all the included salary thresholds (Cattaneo et al. Reference Cattaneo, Keele, Titiunik and Vazquez-Bare2016).

We employ local linear regressions (that is, f ( $\tilde X$ ) is of polynomial order 1) with a triangular kernel fit within the MSE-optimal bandwidth selected by the robust, data-driven procedure by Calonico et al. (Reference Calonico, Matias and Titiunik2014) (our key result is insensitive to bandwidth size, as shown in SA Figure C4). We report the bias-corrected point estimates and standard errors, clustered at the locality level. The key identifying assumption is that the potential outcomes are continuous and smooth across each threshold, implying that: (a) except mayoral salary, no other (pre-determined) municipality characteristics change discontinuously at the cut-off; (b) politicians do not sort their municipalities precisely on one or the other side of a salary threshold. We evaluate these implications in detail in SA Section C.1.4. Observable municipal characteristics, including lagged corruption risk indicators, are broadly balanced, and there is no evidence of potential strategic sorting.Footnote ¹⁰

Europe-Wide Results

Column 1 of Table 1 shows the results of our fixed effects (FE) model from Equation 1. Using the full sample of contracts and municipalities, a 1 per cent increase in mayoral salary is associated with a precisely estimated decrease in the CRI of 0.019 points, or about 1/13 of the sample standard deviation in CRI. Based on this estimate, an average increase in salary of 15 per cent (panel (a) of Figure 1) is associated with slightly more than a one standard deviation decrease in the CRI. To verify that this association is not idiosyncratically driven by municipalities with very low numbers of tenders, the remaining columns limit the sample to municipalities with a progressively larger minimum number of contracts per year. The estimated association is similar, and even somewhat larger.

Table 1. Higher salaries associated with lower procurement corruption risks

Notes: outcome is the corruption risk index (CRI), described in the text. Model specification is given in Equation 1. Standard errors are clustered by municipality.

We next turn to the RD analyses. To unambiguously ascribe the effect to politicians’ salary, no other policy may change at the same thresholds (Eggers et al. Reference Eggers, Freier, Grembi and Nannicini2018). This is the case for twenty-five of the ninety-seven thresholds in our data, across six of the eleven countries. Henceforth, we call these thresholds ‘unique’.Footnote ¹¹ The remaining thresholds entail compound treatments that include the effect of one or more other policies varying along with salaries, such as the size of the municipal assembly or fiscal transfers from the central government. We therefore face a trade-off: unique thresholds provide greater clarity and internal validity, but limit the sample size and generalizability of an already local RD estimand. Our preferred sample is of the unique thresholds; however, we also report the results for a broader sample of (‘compound’) thresholds with no more than two other policies co-varying (fifty-three thresholds covering ten of the eleven countries), as well as all the thresholds (as in the FE analysis). SA Section A indicates which threshold belongs to each sample.

Panel (b) of Figure 1 shows descriptively the change in the CRI at the cut-off across the unique thresholds. Similar to the correlations from the FE model, there is a clearly identifiable drop, from around 0.3 (on a 0–1 scale) below the salary discontinuity to 0.18 immediately above it. That said, there is a fair bit of variability in the CRI, especially above the cut-off (as indicated by the dots, which summarize in small ‘bins’ of the running variables the unplotted raw values of CRI for each contract). In Table 2, we therefore report the RD effects using the model in Equation 2. The first column of Table 2 shows the effect across the unique thresholds.Footnote ¹² At the cut-off, the CRI decreases by 0.078 points (p < 0.01), from about 0.12 to 0.04 (on a 0–1 scale) – a 67 per cent decrease.Footnote ¹³ Since this estimate draws on a very limited sample of contracts (from 165 municipalities, or about 1per cent of municipalities in our data), column 2 uses the sample of compound thresholds that includes slightly more than half of all thresholds.Footnote ¹⁴ The estimated effect drops by roughly 35 per cent but remains precisely estimated (−0.051, or about a 20 per cent decrease, p < 0.01). Column 3 reports the effect using the entire sample, with the caveat that the confounding may be substantial given that (many) other policies vary alongside mayoral salaries. The point estimate is smaller and less precisely estimated, but still suggests a drop in the CRI at the cut-off of 0.02 points, or about 5 per cent. Altogether, these results point to a plausible negative effect of higher salaries on procurement corruption risk, ranging in magnitude from relatively small to relatively large.

Table 2. Higher salaries reduce procurement corruption risks

Notes: outcome is the corruption risk index (CRI), described in the text. Model specification is given in Equation 2. Standard errors are clustered by municipality. RD: regression discontinuity.

These negative effects may be driven not by the decrease in corruption risks but a change in the composition of contracts that may impact our red flag indicators. For example, an average tender in higher-salary municipalities may be of higher value, which in turn may attract fewer bidders (for example, because they involve more complex products or services), and fewer bids may justify shorter submission and decision periods, or result in greater spending concentration. The RD design, however, should rule out these compositional differences. SA Table C6 confirms this, as we observe no salary effect on average contract value, contract volume, or average procurement expenditures.

To further assess the plausibility of our results, we evaluate three additional implications. First, if the effects are indeed driven by the sensitivity of corruption to wages and not some other aspects of municipalities correlated with changes in salaries, we should observe stronger effects at thresholds with larger salary increases. SA Table C7 confirms this, as the negative effect is larger at the larger-increase thresholds. Second, if CRI captures the risk of corruption rather than some other characteristic of contracts, markets, or municipalities, we should observe a stronger effect where corruption is a greater concern. SA Table C8 confirms

this, as the negative salary effect is concentrated at thresholds in municipalities located in more corrupt EU regions. Third, since corruption increases waste in public spending, we should observe that higher salaries increase the efficiency in procurement. CA Table C10 (column 1) confirms this: at the cut-off, the average ratio of the final bid price to the ex-ante estimated price falls by about 5 percentage points.

Evidence from Difference-in-Discontinuities in Romania

To further alleviate concerns about the influence of unobserved differences across municipalities (even in the vicinity of salary thresholds), we employ an alternative design that exploits changes in the location of a number of population-based salary thresholds in Romania. In 2018, six new salary thresholds were introduced and four existing thresholds abolished, with three thresholds remaining unchanged (SA Section C.2.1 gives more details). This set-up allows for a difference-in-discontinuities design, where we can compare trends in the CRI close to the salary thresholds among two groups: localities affected (treatment group) and unaffected (control group) by the 2018 reform.Footnote ¹⁵ SA Section C.2.2 shows that the key (additional) identifying assumption of parallel trends is plausible, and that pre-determined locality characteristics in both treatment and control groups are well-balanced.

If higher salaries indeed lead to lower procurement corruption, we should observe a persistent negative RD effect in the control group, but a change in the effect in the treatment group in line with the change in the salary thresholds. Patterns in Table 3 are consistent with these expectations. We run RD models similar to that in Equation 2.Footnote ¹⁶ Standard errors are again clustered by locality. Column 1 reports the placebo salary effect: it compares the CRI in localities around the cut-off in the treatment group either before the salary thresholds have been established or after they have been abolished. As we should expect, the RD effect is essentially zero, both substantively and statistically. By contrast, in column 2 the effect

Table 3. Similar salary effects from difference in discontinuities in Romania

in the presence of actual salary thresholds (after their introduction or before their abolition) is much larger and statistically more precise, with CRI decreasing by 0.2 points, or about two-thirds (p < 0.05).Footnote ¹⁷ Columns 3–6 separate these effects by whether the thresholds were removed or introduced. In both cases, the patterns are consistent with expectations. In localities where the new thresholds were introduced in 2018, we see no difference in the CRI before the reform but a clear negative effect afterwards (−0.180, p < 0.01). In localities where the thresholds were removed in 2018, we see a large negative effect before the removal (−0.238, p = 0.07) but a significantly smaller and noisier difference thereafter (−0.109, p = 0.49).Footnote ¹⁸

Finally, in columns 7−9 we report the results for the localities in the control group, where the salary thresholds remained unchanged throughout the period. As expected, we observe a negative salary effect both before and after 2018, and the point estimates vary only slightly in magnitude (though somewhat more in precision). In sum, the patterns in Table 3 are consistent with our cross-national results.

Discussion

In this letter, we used comprehensive public contracting data involving local governments in eleven EU countries over more than a decade to analyze objective and validated measures of procurement corruption risk (CRI). We matched more than 2.4 million tenders with almost 100 instances of plausibly exogenously determined variations in local politicians’ salaries across more than 18,000 municipalities. Applying three alternative designs, we consistently find that higher salaries reduce corruption. They also increase efficiency in procurement (SA Table C10). Simple back-of-the-envelope calculations (presented in SA Section C.1.12) suggest that an 18 per cent across-the-board increase in local politicians’ salaries could result in net savings of up to PPP €81 billion over the same period of time as covered by our data (roughly fifteen years). That is almost 4 per cent of the European Union’s annual procurement budget.Footnote ²⁰

We note that our study primarily aimed to provide new inferential evidence rather than to comprehensively arbitrate between existing or generate new explanations (Spirling and Stewart (Reference Spirling and Brandon2025)). For example, while our additional analyses suggest that higher salaries primarily reduce procurement corruption through selection rather than incentives, we leave it to future work to evaluate such questions more thoroughly. That said, our study may be helpful in influencing the relative weights accorded to some of the existing explanations. For example, our findings indicate plausible corruption-reducing effects even for salary jumps that are not very large – on the order of 15–20 per cent – indicating that concerns that salary increases need to be unsustainably large to bear fruit (Besley and McLaren Reference Besley and McLaren1993; Van Rijckeghem and Weder Reference Van Rijckeghem and Weder2001) may be overestimated. We therefore hope our results will spur new theoretical and empirical studies on the impacts of salaries on corruption and other aspects of government performance.