More politicians, more corruption: evidence from Swedish municipalities

The current structure on Swedish municipalities can be traced back to a process of change that started in 1952, and ended with massive municipal amalgamations completed in 1974. Between those years, Sweden’s population increased from 7.2 million to 8.2 million, but the number of municipalities fell from 2498 to 278; the number of elected local representatives fell likewise by 200,000 (Erlingsson et al. 2010). The stated reason underlying these changes was to increase the efficiency of local service provision. Naturally, concerns also were expressed that such efficiency would come at the cost of less democratic participation (Nielsen 2003; Wollmann 2004). The average size of Swedish municipal councils reached a maximum in 1988 (47.85 seats on average) and has been falling steadily since then to 44.1 seats today (after the 2014 election).

Because some municipalities have been split, Sweden currently is (and was at the time of the survey) divided administratively into 290 municipalities and 20 regions. Accounting for 62% of public employment in Sweden in 2014, municipalities are the nation’s largest administrative units both politically and economically. Financed mainly by a proportional income tax of approximately 20%, municipalities are responsible for the provision of schooling, child care, and elder care, leaving only the provision of social insurance, higher education, and defense to the central government. Municipalities also handle welfare provision, zoning issues (including building permits and permission to sell alcohol), culture, and public transport.

Total public municipality consumption accounts for 20% of national GDP, compared to 7% of GDP for the central government. With an average population of 30,000, the mean annual municipality budget is USD 217 million. Roughly three-quarters of this amount is spent on child care, primary and secondary education, and care for the elderly.

To measure the prevalence of corruption problems in Swedish municipalities, Erlingsson et al. (2008a, b) conducted a detailed anonymous web-based survey of local top politicians and high-ranking civil servants in 2008. The survey was developed in 2007 in collaboration with several Swedish corruption experts, relying on both direct and indirect reports of corruption as well as anecdotal vignettes. For the survey, the top four politicians (the chair of the executive board, the vice chair of the executive board, the chair of the municipal council and the chair of the municipal audit department), along with the top three civil servants (the municipal manager, the budget manager and the staff manager) in each of the 290 municipalities, were identified and invited to participate in an anonymous online survey. While the politicians surveyed typically are part of the local councils, the civil servants are not political appointees. The total survey population comprised 2024 individuals, and 1184 (58%) of them completed the survey. Among the 35 questions asked in the survey, six dealt directly with corruption problems. These questions are shown in Table 1, and they all are used in constructing our index of corruption problems (results using the six questions separately are presented in the appendix).

The decision to let the respondents be anonymous was made in order to improve the response rate and the quality of the answers. The drawback is that we do not know who has responded to the survey. Still, by studying the survey responses, we can gain some information about the character of the nonresponses. The dataset includes responses from 287 out of 290 municipalities, and for 247 municipalities we have responses from both politicians and civil servants. The response rate does not have any clear geographical pattern, and no substantial differences are evident depending on the size of the municipalities (geographically or in terms of number of inhabitants). If we compare the response rate by types of municipalities, using the classification of Swedish municipalities that the Swedish Association of Local Authorities and Regions has constructed based on structural parameters such as population, commuting patterns, and economic structure, we find that the response rates vary between 52 and 62%. Furthermore, if we compare the two main groups of respondents, we find that propensities to fill out the questionnaire are about the same—59% among politicians and 57% among civil servants. The difference in response rate by gender is only about one percentage point. Thus, the data we have do not indicate any severe problems related to nonresponse bias.

All questions offered multiple response categories, ranging from “never” to “very often” (questions 1–4) and from “strongly agree” to “strongly disagree” (questions 5 and 6). Figure 1 shows the share of respondents reporting inappropriate behaviors, as well as the share of respondents not answering for each of the six questions in Table 1. Response rates were highest for the first question, which inquired directly about personal experiences; most respondents appeared to reject indications of personal involvement without much hesitation. Response rates were lowest for reports on others having accepted bribes. With respect to actual evidence of corrupt behavior, a remarkable 59.8% of all respondents indicated that others in their community had been offered bribes while in office, and more than half (51%) of respondents indicated that bribes had been accepted. A total of 41.4% suggest that the letting of procurement contracts is not impartial, and 23.3% reported having been threatened during their official duties.

In our baseline analysis, we aggregated the answers into a single measure of corruption using principal component analysis (PCA). The first principal component accounted for 35% of the total variation, with the highest factor loadings on questions 2 and 3, where respondents reported on other politicians being offered or having accepted bribes. We normalized the PCA score into a z-score variable with mean 0 and standard deviation 1. As a robustness test, reported in the appendix, we also created and tested alternative indices without substantial changes in the results.

Corruption reports varied both within and across municipalities, with community-level variation accounting for approximately 25% of the total variation in reported corruption. The substantial variation in reporting across individuals may be interpreted as differences in subjective experiences, but may also reflect heterogeneity in reporting. To use the provided information most efficiently, we relied on individual reports (aggregated into individual-level summary scores) in our main empirical specification. As robustness checks reported in the appendix, we show results for specifications where survey answers were aggregated at the municipal level, and where answers to the six questions were analyzed item by item.

Data on municipal characteristics were combined from a variety of sources. Population size and age structure data were collected by Statistics Sweden and compiled by Johansson (2006). We also compiled information on geographical area, number of council seats, percentage of population with some college education, “close” recent election results and whether the same party had been in power since 1995 from the same data source (Johansson 2006). The reference year for education was 2003; “some college education” was defined as the fraction of the population 25-to-64 years of age with at least 3 years of post-secondary education. The reference year for election outcomes was 2006. We defined elections as “close” if the difference between the two major voting blocs (the four rightwing parties and the three leftwing parties) was less than five percentage points.

We also entered in our models the total municipality budget and total municipality income. The former refers to the expenditures of the municipality and is based on Statistics Sweden’s collection of annual municipality accounts. Municipality income is the regional gross domestic product (GDPR) in 2008, calculated by Statistics Sweden, divided by the population of the municipality.

To control for politicians’ genders, we included a variable that measured the share of women among all local politicians in 2007. The information came from Statistics Sweden’s survey of elected representatives in municipalities and county councils.

We used information on the number of local newspaper offices in the municipalities to capture media coverage of local politics. That information was compiled by Bergh et al. (2013). Furthermore, we also used Statistics Sweden’s information on shares of foreign-born and shares of urban population within the municipalities. Urban population was defined as the share of population that resided in the city center. Finally, the variable “major city” was defined as one of the three largest Swedish municipalities, i.e., Stockholm, Gothenburg and Malmö.

The main empirical model we estimate iswhere C _ij is the (normalized) corruption score of government official i in municipality j, Seats is the number of council seats, X _j and R _i are vectors of municipality and respondent characteristics, and δ _r are regional fixed effects. At the municipality level, we control for total spending, population size (linear and quadratic), the share of population of working age and retirement age (20–64 and 65 and older), municipality income, area (thousands of square km), the share of adults with at least a college education as proxy for the average human capital in each municipality, urban population share, and foreign population share. We also control for factors that have been found to affect corruption in previous studies: the share of women in local elected politicians (Swamy et al. 2001), the number of local newspaper offices (Gentzkow et al. 2004) and local political competition (Shleifer and Vishny 1993).

Though we do not aim to uncover a causal effect of public spending on corruption, we control for public spending because Pettersson-Lidbom (2012) finds that municipalities with larger local councils tend to have smaller public expenditures, which may influence corruption. At the respondent level, we include the following variables collected as part of the corruption survey: sex, type of appointment (politician or bureaucrat) and educational attainment. Because those who are new in local politics are less likely to have experienced corruption problems, we also control for the duration of current appointment.

To overcome endogeneity concerns, we make use of the fact that Swedish law Kommunallagen has constitutional status and stipulates a minimum number of seats for each local council depending on the number of eligible voters residing in the municipality on March 1 the year before the election year. The purpose of the rule is to ensure a high level of proportionality and representation of minority populations. Electoral rolls are produced based on a national population register, which is updated daily. Voter cards are sent to all entitled to vote, and no registration is necessary. The minimum number of seats per municipality is fixed for a term. The required council sizes have been constant since 1979, and are summarized in Table 2.

Municipalities are allowed to increase the number of seats above the minimum requirement, and several have chosen to do so in the past. Given that the law is tied in a non-linear way to population, the law is in practice most relevant when municipalities grow in population, and must increase their council size when passing a threshold. Figure 2 shows the minimum as well as the actual number of seats at the municipality level. It is clear that the law is being followed, and it is noted that the requirements are exceeded by a large number of municipalities (59% overall), while 41% of municipalities currently have exactly the number of seats legally required.

Table 3 shows descriptive statistics for all variables used in the analysis.

Given the cross-sectional nature of the data used in Table 4, we cannot rule out that the positive conditional associations observed are at least partially the result of unobserved confounders, or that that the number of seats may respond to perceived corruption (simultaneity bias). To address these concerns, we report two-stage-least squares (instrumental variable—IV) estimates in Table 5. As discussed above, the instrument used is the required number of seats. The main logic of the IV approach is to use variation in the exposure (council size in this case) driven by an exogenous variable (the instrument). As long as the instrument predicts the independent variable of interest but is orthogonal to the residual of the main equation, IV estimation will yield unbiased estimates of the causal effect of interest. As a first step, we examine the predictive power of the instrument, i.e., assess whether the constitutionally required number of seats predicts the actual number of seats conditional on a full set of control variables. This is indeed the case: in all specifications, the F-statistics for the instrument range between 80 and 118. In column 1, we control for the main variables of interest only, and gradually add a larger set of controls in columns 2–4. The overall results remain very similar to OLS, with a small increase in the estimated impact of additional seats. Our most tightly controlled model (Table 6, column 4) suggests that a 10-seat increase leads to 0.26 standard deviation increase in corruption. The slightly larger effect (0.26 standard deviations versus 0.17 in OLS) suggests that the OLS estimate is biased downwards. A possible mechanism is that less corrupt municipalities opt for larger councils. Another plausible explanation is that the exogenous shifts in council size generated by the law (essentially forcing government size to increase from 30 to 40, or from 40 to 50 members) generates more corruption than a gradual increase in council size observed outside of the minimum requirement. As such, the estimated IV coefficients should be interpreted as a local average treatment effect for legally required increases in council size, which may not apply to more gradual increases in council size. We explore this aspect in greater detail in the following section, where we show estimates based on a regression discontinuity design.

As an alternative identification strategy, the discontinuities in the required number of seats can be used to estimate the effect of council size on corruption using a regression discontinuity design (RDD). For RDDs to yield statistically meaningful estimates, a sufficiently dense distribution of outcomes around the cutoff is needed. As illustrated in Fig. 3, the majority of Swedish municipalities have less than 25,000 voters, and the most suitable cutoff for an RDD design is the increase in the required number of seats at 12,000 voters. Table 6 compares average municipality and politician characteristics around that cutoff. Using a population band of 3000 voters on either side of the cutoff, there are 176 records just below the cutoff and 80 observations just above the cutoff, and these observations do not differ with respect to any of the variables available.¹

One of the most critical assumptions underlying the RDD design is that the assignment variable cannot be modified by subjects. In this case, it is worth emphasizing that municipalities are not involved in the process of deciding where eligible voters vote and that no one has to register in order to be able to vote. Statistics on the number of eligible voters are produced by the Swedish Election Authority, based on the national population register, which is updated daily. To confirm that there is no “gaming” around the 12,000-voter cutoff, we plot population densities in Fig. 3. If politicians or other stakeholders were able to modify population numbers to reach specific thresholds of interest, we should see an increase in the number of municipalities with population sizes just above these specific thresholds. As Fig. 3 shows, this is not the case: the distribution of municipality sizes just above and below the threshold seems to be almost the same, which suggests that population estimates have not been modified.

Figure 4 shows the estimated increase in the number of seats around the cutoff. The number of seats increases almost linearly to the left of the cutoff and then jumps by about two additional seats. The fact that a substantial part of the jump occurs just below the cutoff suggests that many municipalities anticipate that population will increase, and increase the number of seats voluntarily.

Figure 5 shows the main regression discontinuity results for the corruption z-score. Average corruption scores are distributed fairly evenly around zero to the left of the cutoff and have a strong cluster in the 0.3 range just above the cutoff.

As Fig. 5 shows, corruption outcomes are relatively noisy above the cutoff. That suggests that the RDD results are potentially sensitive to different trend specifications. Table 7 shows the results of testing different ranges and trend specifications. While the estimates fluctuate a bit—particularly when the range is expanded—the average increase in the corruption score around the cutoff appears to be about 0.4 standard deviations. Given that the average number of seats increases by about two around the cutoff, this result implies a marginal increase in corruption of about 0.2 standard deviations for each additional seat. These estimates are large compared to OLS estimates, suggesting that the exogenous shift in additional seats at this specific cutoff may trigger a substantial increase in (perceived) corruption problems in the affected municipalities.

In an online appendix, we show results from several additional robustness tests that all confirm the main findings. In summary, we tried creating the corruption index using a binary coding, where each question is coded as 1 if the respondent’s answer indicated any type of problem, regardless of frequency or intensity. We also constructed a categorical index where answers were translated into ordinal scales, so that “never” is coded as 0, “very rarely” as 1, “rarely” as 2, “sometimes” as 3 and “often” as 4, and then summing the scores to create the index. Both alternative codings generated results similar to the baseline findings.

Analyzing the six questions separately reveals significant positive findings for the answers to the questions asking if others in the municipality have been offered money and whether others have accepted bribes. In contrast, respondents’ views on the impartiality of public procurement clearly are not driving the results.

Our results also can be reproduced with municipality level regressions using the median and the maximum level of reported problems in each municipality as dependent variables. Intuitively, using the minimum level of reported problems, no significant results are found.

Finally, we verified that the results for seats still hold if expenditures are excluded from the model. The correlation between the number of seats in local councils and public expenditure per capita is only −0.17, indicating that they are two different dimensions of government size, and that no bias is introduced by including them simultaneously in the regressions.