Violence and Fraud

Legitimate claims on power and support from the international community depend on whether the government adheres to democratic principles. Violations of the electoral process are an obvious breach of these principles, and if detected, entail a risk of jeopardizing both popular domestic and international support. For example, international election observers are an important means of ensuring that elections are free and fair, and have been shown to be effective in suppressing fraud.Footnote 15 However, in countries with ongoing political violence, institutions that promote accountability, such as the media and international and domestic election observers, may face limits on where they can safely operate. Consider the following difference: the small country of Kosovo hosted the stunning number of 22,098 election observers in its 2009/10 municipal election, on average one observer per 30 voters.Footnote 16 In contrast, in the 2009 Afghanistan election there was one observer per 580 voters (10,300 international and domestic observers), a penetration rate of about one twentieth that of Kosovo.Footnote 17 Clearly, however, observers were not uniformly distributed and security concerns prevented them from visiting large parts of the country, which effectively created zones outside international supervision. It is in these areas that we should expect fraud to be more likely to happen.

At the same time, however, the impact of increasing levels of violence on fraud should not be monotonic. In particular, we expect the ability and incentive to commit fraud to decline in high-risk places. If we assume that the candidates’ loyalty networks have a higher tolerance for violence than do election monitors and other institutions for accountability promotion, then we would expect a negative relationship between fraud and violence above a certain threshold. Two mechanisms may explain this. First, for loyalty networks to manipulate the election process effectively, they are required to exert some kind of influence at or around the polling station, for example by bribing election officials. If a deteriorating security situation makes it impossible for agents to develop relationships in contested areas or even to reach the polling station, these operations will be unlikely to take place. At the same time, however, we expect a strategic planning mechanism to operate in the same way. Operating a network of fraud perpetrators entails financial and social costs to candidates, so they will refrain from allocating resources to areas that are likely to see complete interruptions of the election process. Whereas delays or small interruptions of the electoral process may count in the perpetrator's favour as long as the (manipulated) results enter the national tally, high numbers of closed stations effectively render the work of election manipulators useless.Footnote 18 The empirical prediction is, therefore, that fraud should be increasing at first in the level of violence, but decrease again after a certain threshold level of violence is reached. Thus, our first hypothesis is:

As we have mentioned above, we treat violence as an external factor that generates conditions affecting the feasibility of implementing fraud. In other words, in our analysis violence is orthogonal to the election – it does happen in parallel, but is not intended to shift the result in favour of one candidate or party. This sets our theoretical story apart from many other works on election-related violence, where (mostly pre-election) violence serves the purpose of tipping the result in a particular direction.Footnote 19

Spatial Displacement of Fraud

So far, we have formulated predictions about the relationship between violence and fraud in general. We now turn to differences in the operation of loyalty networks between the incumbent and the contender. In countries lacking fully democratic institutions, incumbents often have the opportunity to shape election processes in their favour.Footnote 20 One way to do this is to use the electoral apparatus to manipulate the election outcome. In doing so, incumbents can rely on election officials, whom they themselves or their loyal regional representatives have often appointed. In short, we can assume that the incumbent's loyalty networks perpetrating fraud are tightly linked to, if not overlapping with, the people involved in the election on the ground, which provides plenty of opportunities for the election result to be manipulated in the incumbent's favour.

Competitors running against the incumbent are much less likely to have an official position or to be linked to the electoral commission. Their operation is much more limited, as they try to interfere with the election process from outside. One of the most effective strategies will be the bribing of local officials so that ballot boxes can be tampered with or return sheets manipulated. While less effective, this strategy has the advantage of not being tied to a particular location. If circumstances on the ground do not permit interfering with the election at one place, perpetrators are able to relocate and try manipulating the process at nearby locations. This in turn should be reflected in different patterns of spatial displacement of fraud.Footnote 21 If violence prohibits the operation of the incumbent's perpetrators at one polling centre, no relocation will take place since fraud is carried out by people close to the officials assigned to this location. We should, however, observe spatial displacement of fraud for the competitor: high levels of violence at a particular location should reduce fraud, but high levels of violence nearby should increase fraud as it is displaced. We therefore expect:

In other words, Hypothesis 3 means that local violence should be negatively correlated with fraud in support of the contender, but violence in nearby regions should be positively correlated with fraud in support of the contender.

Background

After the 11 September attacks, allied forces invaded Afghanistan in an attempt to eliminate safe havens for terrorists. The ruling Taliban, having established a brutal regime following the Soviet withdrawal, were ousted from power. In late 2001, an interim government, led by Hamid Karzai, was installed. In 2004, a presidential election was held, and Karzai was elected president of Afghanistan, winning 55 per cent of the votes. There were some reports of electoral manipulation in 2004, but unlike in 2009, no systematic assessment was done and disaggregated data were not publicly released, so it is difficult to gauge the degree of manipulation. Karzai's first term in office came to an end in 2009, leading to the second presidential election in Afghanistan. With increased insurgent activity and the heightened presence of military personnel in the country, this election proved to be much more difficult and the result much more controversial than the first one. One effect was a much lower turnout; whereas in the 2004 election, an estimated 75 per cent of all registered voters went to the polls, this number dropped below 35 per cent in the 2009 election. At the same time as the presidential election, the provincial council election was held, but our focus here is on the former.

Afghanistan's constitution requires that the president be elected every five years and requires a simple majority of the votes in the first round.Footnote 22 If no candidate reaches this threshold, a second round must be conducted. The 2009 election was the first election conducted solely under Afghan oversight. The country's first elections in 2004 (presidential) and 2005 (parliamentary) were organized by the Joint Electoral Management Body (JEMB), a collaboration between Afghan officials and international (mostly United Nations) personnel. Before the 2009 election, the JEMB was dissolved, and the Afghan government established the IEC. The IEC has seven members appointed by the president, thus raising doubts about the commission's independence. In addition to their headquarters in Kabul, the IEC has provincial offices in each of the country's thirty-four provinces.

The IEC was responsible for the organization of the entire election process, including voter registration, election material distribution and vote counting. Voter registration was plagued by problems, including multiple registration; as a result, some regions reported numbers of registered votes that exceeded the population eligible to vote. On election day, voting took place at polling centres typically located in public buildings, schools or mosques. Polling centres included one or more polling stations, each of which corresponded essentially to a single voting booth and ballot box. Stations were set up such that each of them could accommodate a maximum of 600 voters. After voters had cast their ballot, their registration card was punched and one of their fingers marked with indelible ink in order to prevent multiple voting.

The ongoing insurgency shaped many aspects of the electoral process. The National Democratic Institute reports that in many areas (especially in the south and south-west), Taliban violence led to reduced turnout, or even prevented polling stations from opening.Footnote 23 Importantly for our purpose, however, is the fact that Taliban violence does not seem to have been perpetrated to affect the result of the election in favour of either of the candidates, but rather to interrupt the process in general. Various media accounts attest to this.Footnote 24 Also, a statistical analysis finds no evidence for an effect of attacks on results for either candidate. Bill Greer, an analyst for GeoIQ, finds no correlations between Pashtun or Tajik population shares (which are strong predictors of electoral behaviour) and violence on election day. He concludes: ‘The attacks taking place on election day were not for or against a specific candidate but more specifically towards the political process and the people of Afghanistan who would support a free and open democratic proces.’Footnote 25 We provide a similar test in the online appendix, which shows that violence around election day is uncorrelated with ethno-linguistic fractionalization at the district level. If violence was perpetrated to favour a particular candidate, we should see violence increasing in fractionalization, as electoral competition goes up.

Initial reports of fraud were widespread. Especially after the first announcement of results was made on 16 September, numerous problems in the results became apparent (for example, the high number of polling stations reporting the maximum turnout of 600 or higher). Rising international pressure led the Electoral Complaints Commission (ECC) to order a physical examination of a sample of ballot boxes, which confirmed a high number of instances of ballot box stuffing. Clearly, the 2009 Afghanistan election was severely manipulated, and no sophisticated test is required to confirm this. However, as we will show below, fraud displayed great regional variation, and we can exploit this variation to learn more about the implementation of electoral manipulation.

The vote-counting process worked as follows. Every polling centre was assigned a polling-centre manager. The polling-centre manager was responsible for transferring all ballot papers as well as a tally sheet for votes cast in the polling centre in a sealed Tamper Evident Bag (TEB) to the District Field Co-ordinator (DFC). Polling-centre officials were responsible for filling out the tally sheet. Tally sheets should have been sealed in TEBs and then placed into a designated ballot box with all supporting materials. No TEBs should have been opened until they reached the tally centre in Kabul. Tally sheets in Tamper Evident Bags were transferred directly from the District Warehouse to the National Tally Centre in Kabul with no intermediate stop at the provincial level. Ballot boxes remained in District Warehouses, the only boxes that were ever sent to the National Tally Centre were those that were summoned due to complaints or during the random sampling of boxes for the recount. Many of the TEBs showed signs of manipulation, but there was no transparent process at the Tally Centre for setting aside suspicious tally sheets. Obviously, there were different entry points for (election-day and post-election-day) fraud in Afghanistan. However, due to the responsibility given to the polling-centre manager and the DFC, there is reason to assume that most of the electoral manipulation had been committed before a return sheet reached the National Tally Centre in Kabul.

Unlike many other instances of election manipulation, fraud had a distinctly local character in Afghanistan. Even though we lack systematic evidence for the 2009 election, subsequent work documents extensive fraud of exactly this type in the 2010 parliamentary election in Afghanistan.Footnote 26 A sample of 471 polling centres across nineteen provincial centres, has yielded primary reports of candidates’ agents illegally damaging or removing polling materials in about 19 per cent of the sample and that in at least 339 of the 471 polling centres indelible ink used to prevent multiple voting was absent or could easily be washed off. Additionally, polling centres in the sample were the subject of 0.6 formal complaints each to the ECC specifically about the polling-centre manager, with a much larger number of general reports made to the ECC. While, to our knowledge, there are no estimates of the relative intensity of polling centre fraud relative to other strategies, this suggests our assumption of the local nature of fraud is empirically supported.

With Afghanistan's long history of patronage politics, there is reason to assume that local loyalty networks were involved in shaping the election outcome. Afghanistan has long been dominated by patronage along ethnic/tribal lines, and recent attempts to move beyond this were largely unsuccessful. Some authors even argue that the Bonn Agreement in 2001 reinforced, rather than avoided, intra-government competition between ethnic factions.Footnote 27 Thus, the young Afghan state that we study in this article has moved towards what is called a ‘patronage-democracy’ – a state where the allocation of government jobs and services is done with considerable discretion, and can thus favour particular societal groups.Footnote 28 In contemporary Afghanistan, this practice is commonplace. For example, not only can the president appoint province and district governors, but also the chiefs of police and, most importantly for our purpose, the officials of the electoral commission.Footnote 29 These selective benefits and rewards give rise to the creation of ethnic/tribal ‘loyalty networks’ – networks of support for particular political candidates that obtain these benefits, but in return need to pledge support for their candidate's political campaign. While these loyalty networks align primarily with the candidates’ ethnic affiliation (Karzai: Pashtu/Popalzai tribe; Abdullah: Tajik), other factions (Hazaras, Uzbeks) have struck deals with candidates.Footnote 30 So, ethnic and political divisions did not align perfectly, but mobilization and political support was predominantly organized along the former.

Measuring Electoral Manipulation

Election fraud is conducted secretly and so is difficult to measure empirically. Most existing sources of data on election fraud rely on information from immediate observers of electoral manipulations.Footnote 31 For example, ethnographic interviews can be used to find out about the context of particular violations of electoral integrity. Also, complaints filed by victims of electoral misconduct can serve as data sources for the study of fraud, as for example in Ziblatt's analysis of the German elections.Footnote 32 While these data sources can help us find out about the type of manipulations and the context under which they occurred in particular instances, they are less suited for an analysis of the general pattern of fraud. First, as some authors mention, they are not objective and may suffer from various problems, including partisan bias.Footnote 33 Secondly, individual reports are likely to be incomplete: we cannot infer from the absence of reports at a particular location that there was no fraud; rather, something could have prevented people from observing and reporting manipulations.

Recognizing the different weaknesses inherent in incident reports, researchers have tried to develop alternative ways to measure fraud. These new ‘forensic’ techniques explore deviations from standard numerical patterns in vote counts, turnout and party/candidate shares. The volume by Myagkov et al. provides an excellent discussion of these techniques in the context of elections in Russia and Ukraine.Footnote 34 Almost all of the techniques presented exploit relationships between different election-related results, for instance the relationship between turnout and a candidate's vote share. However, the data availability for the 2009 Afghanistan election is very limited; because of the problematic voter registration procedure (disaggregated results of which were not released publicly), the only information available to us are the election results at the polling-station level. Given these limitations, we resort to a different forensic technique that detects deviations in digit patterns in election results.

The most frequently used test of this kind is based on Benford's law.Footnote 35 However, one problem with this test is that a variety of fraud mechanisms can generate the deviations it detects, which leaves considerable uncertainty about the type of manipulation that occurred.Footnote 36 Also, the polling-station level turnout numbers we use are fairly low (capped at 600, frequently below 100), which creates problems for the Benford test. For this reason, we resort to another type of digit pattern-based fraud test, which we introduce in the next section.

The Last-Digit Measure

Recent work by Beber and Scacco develops a novel set of forensic fraud measures.Footnote 37 The assumption is that rather than describing true election outcomes, voting results in manipulated elections are made up by people attempting to make these numbers look random. In generating these seemingly random numbers, however, people follow certain patterns which deviate from true randomness and, therefore, can be detected by statistical tests. For example, when reporting election results, individuals often choose to plug in rounded numbers that end in 0 or 5. A simple test that examines the distribution of last digits in election results is able to detect these deviations.Footnote 38 In a fair election, the last digit recorded at any polling station x_i should be drawn from a uniform distribution, X : U(0,9). It can be proven formally that this should be so using minimal distributional assumptions on the underlying data-generating process.Footnote 39 Deviations from the uniform can then be tested using a chi-square test. If this test detects a significant deviation from the uniform, we have evidence that the results are not clean.

A clear advantage of this test is the limited data requirement: all we need are election results at a disaggregated level (the polling station), and no other information (such as turnout) is required. Another upside is the potential for disaggregation, similar to other digit-based fraud measures.Footnote 40 In order to exploit variation in levels of violence and fraud, we require measures of fraud that provide information on the degree of ballot manipulation at the district level, because this is the finest level of resolution that can be analysed with our data.Footnote 41 The last-digit test can easily be disaggregated to the district level, by simply running a chi-square test on the last digits for the polling stations in each district. In addition to the overall pattern of fraud across districts, we can also disaggregate voting totals by candidate and analyse whether the votes reported for a candidate show signs of manipulation. This is the approach taken in our spatial displacement tests presented below.

To develop intuition for the last-digit method, Figure 1 graphs the distribution of final tally digits for the total vote count in three districts from different regions in Afghanistan. According to the last-digit test, there is clear evidence of fraud in two of these districts. In Du Ab (Nuristan Province, ISAF Regional Command East) and Arghandab (Kandahar Province, ISAF Regional Command South), we see clear departures from the uniform. The p-values reported upper right confirm that in Du Ab we can reject the uniform digit null at the 99 per cent confidence level and in Arghandab we can reject at the 95 per cent confidence level. In our empirical analysis below, we use the district-level p-values to measure fraud.

Fig. 1

Measuring fraud using the Beber–Scacco last-digit test in three Afghan districts Notes: The panel shows the distribution of last digits across the vote totals reported by the stations in the respective district. Du Ab and Arghandab show suspicious deviations from the uniform distribution (an inflation of zeros), which are picked up by the chi-square test.

What does the last-digit measure capture? Fraud detection indicators based on Benford's law have recently been criticized for failing to identify manipulation correctly.Footnote 42 The general problem is that deviations picked up by these indicators can, but need not, be due to fraudulent manipulation. Moreover, fraud can occur without being reflected in the digit patterns. In short, there is an inherent risk of both Type I and Type II errors in the use of digit-based fraud measures. There are two ways in which we can address these criticism in our case study.

First, we can make sure that the type of digit-based test corresponds to an empirical mechanism of fraud. The underlying motivation for the last-digit test is about people filling in return sheets with false vote counts, which is a mechanism with a relatively clear ‘local’ focus.Footnote 43 We believe that this type of manipulation occurred in Afghanistan. At the polling centre, the manager had unsupervised access to ballot boxes and was the single responsible person for filling in the return sheet and sending it back to the district. In the district centre, much of the oversight was centralized in the hands of the DFC. Indeed, this assumption is consistent with the observation that many return sheets contained obviously fraudulent numbers, but monitors at the corresponding polling centre reported no fraud. In sum, the last-digit fraud measure may be suitable to capture much of the election fraud in this particular election, which occurred somewhere between the polling centre and the district. At the same time, however, some types of local election manipulation will not be captured by this measure. For example, attempts to intimidate or buy off voters would be unlikely to lead to deviations in the last-digit pattern. Similarly, the last-digit test would not be sensitive to instances of multiple voting by the same person. These caveats need to be kept in mind, as we are working with an indirect way of observing fraud that necessarily comes with some uncertainty.

The second way to address criticism against the use of the last-digit measure is to validate it against other evidence of fraud. Frequently, however, this will not be possible, since we do not have this evidence: As discussed above, reports of individual instances of fraud are insufficient to perform such a validation, because they lack completeness and are potentially biased. To our knowledge, there have been few validation attempts of the last-digit measure. In the case of the Afghanistan election in 2009, however, a detailed examination and recount of the ballots after the election offers a rare opportunity to do this; the next section introduces a fraud measure based on evidence from the recount.

Validating the Last-Digit Measure

On 8 September 2009, the Electoral Complaints Commission (ECC) – an independent body established with United Nations assistance to adjudicate challenges and complaints related to the electoral process – issued an order to the IEC to audit and recount votes cast in polling stations that fell into one of five categories.Footnote 44 From each category, a random 10 per cent sample was taken (342 boxes in total), each of which was subsequently opened and examined for physical signs of fraud.Footnote 45 As a result of this examination, the IEC reported the share of ballot boxes in each of the categories that were shown to be fraudulent. More information about the recount can be found in the online appendix.

We begin by illustrating the computation of this measure for three districts presented in Figure 2. The figure graphs the share of polling stations in each of the three districts that satisfy each of the five potential fraud categories. The dark shaded portion of each bar corresponds to the share of votes found to be actually fraudulent at the national level for each of the categories. Our measure multiplies the share of polling stations that fall into each of the suspicious categories in a given district by the estimated national share for that category. Graphically, our district-level measure of the share of fraudulent boxes is then the sum of the dark shaded portions.Footnote 46

Fig. 2

Estimating fraud shares in three selected provinces with data from the national recount Notes: The bars show the proportion of polling stations in the respective category. The dark shaded part is the share of polling stations found to be fraudulent in the respective category.

We can now compare our last-digit measure of fraud to the recount-based one. Figure 3 plots our recount-based measure against the p-value obtained from a chi-square test for each of the 389 districts in our sample. We overlay a kernel regression which provides the non-parametric prediction of the mean for the recount-based measure conditional on the p-value of the last-digit test. We observe that, beginning at about 20 per cent significance, we see a clear positive relationship. According to the prediction, if we reject the null of no fraud at 95 per cent confidence – the level we use for our analysis – we estimate that roughly 37.5 per cent of boxes in the district are manipulated. Moreover, we can see that almost all of the districts with a high share of manipulated boxes (0.6 and above) are significant at the 0.1 level in the last-digit chi-square test. An additional step confirms this: in a regression of recount-based fraud share on last-digit p-value in all districts, the latter obtains a strongly significant positive coefficient (see online Appendix for detailed results).

Fig. 3

Scatter plot of the last-digit and the recount-based fraud measure at the district level Note: X-axis: p-value of last-digit chi-square test. Y-axis: share of manipulated boxes according to the national recount.

In sum, these results provide clear evidence that the Beber–Scacco measure applied at the district level successfully predicts a great deal of the tally manipulation that took place during Afghanistan's 2009 election. Beyond our analysis, however, we believe that this measure may be useful in other settings. Given the trend towards increased reporting of results at the polling station level in democratizing countries, the last-digit test can be used to produce reliable estimates of fraud even in cases where other information about electoral manipulation is difficult to obtain.

The Dependent Variable: Fraud

Our fraud data are computed from polling-station level data made public by the International Election Commission (IEC) on 19 September 2009.Footnote 47 As described above, we group the polling stations by district and apply the Beber–Scacco last-digit test to the total vote count. We code a binary dependent variable for fraud, which takes the value of 1 if this test is significant at the 5 per cent level for a particular district. For our spatial displacement models, we apply this test separately to the vote counts for the incumbent (President Hamid Karzai) and the strongest contender (Abdullah Abdullah). In order to exclude the possibility that our results are driven by the number of closed polling stations (which may distort the last-digit test results), we include the number of closed stations at the district level as a control variable.

Independent Variables

Population

High quality population data for Afghanistan are not available, as the last census was carried out in 1979 and remained incomplete. We rely on the LandScan population data, raster GIS datasets that represent population figures at the level of small cells.Footnote 51 The LandScan data estimate the population for 30” × 30” latitude/longitude cells across the globe. Population counts are apportioned to each grid cell based on an algorithm that takes into account proximity to roads, slope, land cover, night-time illumination and other information available from Landsat satellite observation. These estimates are aggregated to the district level for our analysis. Summary statistics for all variables are given in Table 1.

Table 1

Summary Statistics for the Variables Included in the Regression Analysis