1 Introduction

The literature on legal accountability is playing catch-up to the literature on decision-making in judicial review cases. Relatively little is known about the factors that influence judges as they consider imposing civil and criminal penalties, and even less in countries outside North America and Europe. The gap is especially pronounced with corruption: despite growing numbers of cases around the globe, the literature on judicial decision-making in corruption prosecutions remains sparse. Why do judges convict some defendants in corruption cases and not others? Why do judges apply harsher sentences to some defendants than others? In sum, what explains variation in the severity with which the judicial system judges public figures and private citizens accused of corruption?

These questions are clearly relevant both for public debate and academic theory, but have only recently been raised in the specialized literature. One cause of this neglect may be data availability: there have not always been sufficient numbers of cases to evaluate judicial behavior regarding legal accountability. Associated with this is a problem of comparison: seldom are cases comparable to one another along a range of dimensions, holding constant legal system, jurisdiction, statutes, case details, or contextual background. One consequence is that most academic research of legal accountability in corruption cases tends toward qualitative analyses of aggregate patterns, using longitudinal case studies or small-N comparative research.

Brazil’s Operation Car Wash (Lava Jato, in Portuguese), perhaps the “the largest corruption scandal ever to beset at democratic nation” (Fisman & Golden, 2017, 13), offers a unique opportunity to address these questions, partly because of the large number of cases that were heard in a single courtroom, by a single judge.

Operation Car Wash was a criminal investigation into grand corruption and money laundering that began in March 2014 in the Brazilian city of Curitiba. It began as an inquiry into irregularities at the state-controlled oil firm Petrobras, Brazil’s largest company. Investigations progressed rapidly, unveiling evidence of corruption totaling billions of dollars, covering a wide variety of government agencies, and implicating politicians, bureaucrats, businesspeople and financial operators. Over the course of nearly seven years, the operation indicted, convicted and imprisoned hundreds of individuals (Rodrigues, 2020; Lagunes & Svejnar, 2020; Da Ros & Taylor, 2022). As Operation Car Wash both fueled and was fueled by severe political and economic crises that brewed simultaneously in Brazil over the 2010s, many studies cast an unflattering light on the investigation’s consequences for Brazilian democracy (Sá e Silva, 2020; Kerche & Marona, 2022; Santos & Gallego, 2022).

At the heart of the operation was the 13th Federal Trial Court of Curitiba, specialized in money laundering crimes. It was presided over by Judge Sérgio Moro between March 2014 and November 2018, when Moro resigned from the judiciary to become Minister of Justice. From Curitiba, the operation spread to several courts around the country and even beyond Brazil. But it was in Curitiba that the most consequential decisions were rendered, including the convictions of leading politicians such as the former president of the Chamber of Deputies Eduardo Cunha, former presidential chief of staff José Dirceu, former Finance Minister Antônio Palocci and former President Luiz Inácio Lula da Silva. Prominent business people, long accustomed to impunity, were also convicted, including the scion of the Odebrecht construction empire, and many of his ostensible competitors.

The prominence of the defendants, the political repercussions of the case, Judge Moro’s own career choices after he left the judiciary, and the intense public awareness of Operation Car Wash have meant that the trial court’s decisions are frequently interpreted through a subjective lens, defined by observers’ political leanings or their perspectives on how corruption should be addressed. Operation Car Wash, Moro, and Moro’s decisions have increasingly become an object of criticism over the last few years. Many of Moro’s decisions would later be overturned by the Brazilian Supreme Court—notably the conviction of former president Luiz Inácio Lula da Silva—due to different allegations of abuse and bias while Moro presided over the investigation (Da Ros & Taylor, 2022; Prado & Machado, 2022a, b).

Despite these important caveats, we believe that analyzing Moro’s decisions on their own—without reference to his later career decisions or the subsequent decisions of higher courts—is an important exercise for understanding the factors that determine judicial decision-making in corruption trials. Seldom has a trial court judge in any jurisdiction presided over such a large number of interconnected corruption cases. The extensive dataset of cases compiled from Moro’s decisions, moreover, enables this paper to take an arms’ length perspective, analyzing Moro’s sentences through multivariate quantitative analysis of the effect of two sets of variables on judicial decisions regarding conviction and sentence length: variables related to defendant profiles and to prosecutorial strategies. In so doing, we hope to contribute both to the broader global literature on the judicial role in anticorruption, as well as to the debate within Brazil concerning Judge Moro’s decisions and the workings of Operation Car Wash, building upon the predominantly qualitative literature on these topics.

To that end, the next section offers a conceptual framework and lays out the hypotheses that guide the analysis. The third section details the data and method. The fourth presents the results. The paper concludes with a summary of the findings and suggestions for future research.

2 Legal Accountability: Definitions and Hypotheses

Questions regarding variation in the incidence and intensity of judicial responses to corruption are closely related to the debate on accountability. Accountability refers to the obligation of public officials to explain their actions and to answer for them. Accountability has been classified in a variety of types, which seldom explicitly reference the judiciary (Schedler, 1999; Manin et al., 1999; Taylor, 2019). For the purposes of this paper, we refer to the specific judicial role in accountability as legal accountability, defined as the enforcement of civil or criminal sanctions by judicial agents (Bovens, 2007; Lindberg, 2013; Da Ros, 2019).Footnote 1

The academic literature on judicial behavior has focused predominantly on supreme courts and constitutional review, seeking to understand which factors—political, social, institutional etc.—influence the decisions made by judges and justices in the exercise of their duties (Da Ros & Taylor, 2019; Hilbink & Ingram, 2019). The vast comparative literature on judicial behavior has used judicial review cases to theorize a variety of explanatory models, such as the attitudinal, strategic and institutionalist.Footnote 2

Theorization of how judges decide corruption cases, however, remains incipient. There has been almost no adaptation of theories from other areas of judicial behavior to corruption cases, nor has there been much theorizing derived from corruption cases themselves. In part, this is because of a methodological bias that hinders both the formulation of new theories and the testing of existing theory: most approaches to legal accountability in corruption rely on aggregate data (e.g., the total number of criminal convictions per country or period). The unit of analysis is often the entire collectivity: cases in a given country or state, for example. As a consequence, conclusions are drawn by aggregation: patterns of corruption and accountability in a given jurisdiction. This approach does not allow for analyses in which each judicial decision constitutes the unit of analysis (King, 1997).

The most typical aggregate approach is the longitudinal case study, seeking to explain why the judiciary in a particular country has come to convict a larger or smaller number of defendants for corruption over time. Generally motivated by the occurrence of major investigations that profoundly impacted the political system, the effort of these studies is to understand why prosecutors and judges began to behave more assertively. Examples of this approach include a variety of exemplary case studies, such as analyses of the Mani Pulite investigations in Italy in the 1990s and its consequences over the following decades (della Porta, 2001; della Porta & Vannucci, 2007; Vannucci, 2009; Dallara, 2019; Manzi, 2021), the corruption trials in France during the 1990s (Roussel, 1998; Garraud, 2001; Adut, 2004), recent prominent cases in Latin America (Conagham, 2012; Michener & Pereira, 2016) and, of course, studies of Operation Car Wash itself (Lagunes & Svejnar, 2020; Rodrigues, 2020; Sá e Silva, 2020; Kerche & Marona, 2022; Da Ros & Taylor, 2022).

A second approach also relies on aggregate data, but does so comparatively. Such studies seek to understand why some polities have produced more legal accountability (i.e., investigations, prosecutions, convictions) than others. Such studies typically draw from a relatively small number of cases, dedicating themselves to comparing different countries, states or governments in Western Europe (Sousa, 2002; Maravall, 2003; Sims, 2011; Guarnieri et al., 2020), Latin America (Da Ros, 2014; Ang, 2017; González-Ocantos & Hidalgo, 2019; Pimenta & Greene, 2020) and Eastern Europe (Popova & Post, 2018).

One of the consequences of the predominance of aggregate approaches is that the use of qualitative methods prevails. Multivariate quantitative studies based on individual data have been less common, despite recent efforts in this direction (Ang, 2017; Gehrke, 2019; Bento et al., 2020; Mancuso et al., 2021). This is because aggregate approaches often reduce the number of units of analysis by treating the available data in aggregate form. That is, because they are concerned with explaining the overall pattern of charges or convictions, these studies often do not inquire into which factors influence individual decisions by prosecutors and judges to charge or convict defendants, or which factors impact the penalties imposed on convicted defendants.

This paper disaggregates the data, using each of the individual judicial decisions against specific defendants as units of analysis. The use of individual decisions permits us to employ multivariate techniques that complement existing qualitative perspectives. Our hypotheses are derived from aggregate research on the subject and from studies on judicial behavior in judicial review cases (an area of research in which individual approaches predominate).Footnote 3 The hypotheses fit into two pairs, regarding: 1) defendant characteristics; and 2) prosecutorial strategies.

The first pair seeks to understand judicial decisions to convict, and the penalties imposed, based on defendant characteristics, such as profession and, in the case of politicians, their party affiliation. With regard to profession, our hypothesis is that judges may act differently depending on whether they are judging politicians, businessmen, bureaucrats, money launderers or unwitting accessories (laranjas). Politicians are expected to be less likely to be convicted, in line with the strategic approach to judicial behavior, because courts are likely to be reticent to move against those who can wield power against them (Maltzman et al., 1999). Indeed, most judicial anticorruption efforts seemed targeted at petty corruption in Brazil until recently (Madeira & Geliski, 2019; Levcovitz, 2020). We assume that more powerful defendants, or those with greater capacity to damage the judiciary or individual judges, are likely to be treated more leniently than less powerful defendants (H1a). The second hypothesis is that judges may decide cases differently depending on defendants’ political affiliation (H1b). We test whether there is differential treatment of politicians by party. This hypothesis conceives of judicial behavior as an extension of party competition (Gordon, 2009; Ríos-Figueroa, 2012; Ang, 2017; Popova & Post, 2018), and is closely aligned with the attitudinal model, in which judges’ political attitudes are considered major motivators of judicial behavior (Segal & Spaeth, 2002).Footnote 4 This produces two hypotheses:

H1a:

The defendant’s profession impacts both the likelihood of convictions and the length of judicial sentences; and

H1b:

The defendant’s political party affiliation impacts both the likelihood of convictions and the length of judicial sentences.

The second pair of hypotheses refers to prosecutorial strategies. Even though the cases all involved corruption, defendants faced diverse charges. Brazilian law dictates different sentences for different crimes, ranging between 2 and 12 years for active and passive corruption (roughly, bribe-paying and bribe-receipt), between 3 and 10 years for money laundering, and between 3 and 8 years for criminal organization, for instance. Defendants may also face more severe charges, for example, for higher volumes of money laundering. The first hypothesis of this second pair (H2a) is that some types of crime are likely to lead to more severe treatment, and the higher the number of crimes for which defendants have been charged, the more likely this is to lead to conviction and heavier sentences. The second hypothesis (H2b) is that the type and quantity of evidence will increase the severity of sentencing. Evidence may come from a variety of sources, including prosecutorial witnesses, phone intercepts and electronic data (including text messages, messaging apps, email, and cloud data), audit reports, and information provided by the financial intelligence unit COAF and the Federal Revenue Service, among others. This information allows us to evaluate how the information prosecutors provide influences judges’ behavior (Machado & Paschoal, 2016; Arantes & Moreira, 2019). We also take into account plea bargain agreements, which likely affect plea bargainers’ chances of conviction and the severity of sentencing and other defendants tried within the same case. The two hypotheses are expressed as follows:

H2a:

The types and the quantity of crimes defendants have been charged with impact both the likelihood of convictions and the length of judicial sentences; and

H2b:

The types and the quantity of evidence presented in court impact both the likelihood of convictions and the length of judicial sentences.

3 Data and Descriptive Analysis

The dataset used to analyze these hypotheses was compiled from the website maintained by the Federal Judiciary of Paraná (JFPR), which encompasses the 13th Federal Court of Curitiba and permits full consultation of all criminal cases tried by Judge Sérgio Moro between March 2014 and November 2018. Similar data has been used previously by other scholars, such as Fontoura (2019) and Rodrigues (2020); we compiled the dataset used here directly from the JFPR website, checking it against the data compiled by those two authors.Footnote 5 We have downloaded the sentences and decisions on interlocutory appeals (embargos de declaração, which are appeals within the same court, akin to requests for amendment of judgement). Having analyzed the text of the decisions, we compiled the database for the variables of interest.

This study does not cover the totality of Sérgio Moro′s behavior within Operation Car Wash. We only analyze his sentences quantitatively. This means that the paper ignores both the judge’s off-the-bench behavior, such as interviews, press releases and opinions on legislative proposals (Hilbink & Ingram, 2019). We also leave aside Moro’s control over the court’s agenda, the timing with which cases were dispatched, and several consequential decisions taken outside the strict limits of the cases, such as the decision to publicize recorded audios of conversations between former President Lula and then President Dilma Rousseff. Similarly, our research does not cover all sentences offered under the umbrella of Operation Car Wash. Other judges also handed down decisions, including Moro’s successors at the 13th Federal Court of Curitiba after 2018, as well as judges participating in related cases in Rio de Janeiro, São Paulo and Brasília, including higher courts.

That said, Moro’s role was significant, setting both the tone and the tenor for Car Wash, and representing the bulk of the decisions made in the investigation from 2014 to 2018. During almost the entire period he ran the 13th Federal Court of Curitiba, Moro’s court had unique jurisdiction over cases in Operation Car Wash, judging a total of 46 criminal cases filed against 193 different defendants on the merits. Because some defendants appeared in more than one case, Moro actually made 289 decisions. These decisions constitute our units of analysis. We exclude those cases in which the sentence was suspended by the judge because the defendant signed a plea bargain agreement and had already served the sentence established in the agreement (e.g., Alberto Youssef in case n. 5028608-95.2015.4.04.7000), as well as those cases in which the defendant died before trial (e.g., Humberto Mesquita in case n. 5025676-71.2014.4.04.7000). Finally, we excluded cases of defendants whose indictments were not accepted by the court (denúncia não recebida) so that they could be tried in other related cases.Footnote 6

Because we are interested in what factors lead to conviction and the length of the penalties imposed, the dependent variable was measured in two ways. The first was through a dummy variable for conviction. A zero score might occur either because of acquittal or because the defendant “ceased to be convicted” on account of a lis pendens situation (of parallel proceedings; e.g., Waldomiro de Oliveira in case n. 5083258-29.2014.4.04.7000).Footnote 7 The second measure of the dependent variable is a continuous measure of the length of sentence for convicted defendants, ranging from 18 to 250 months.

For defendants who signed plea bargain agreements, we use the sentence length stipulated in judgment before the application of the agreements’ benefits. We take this approach because most agreements unified the penalties of the defendants not only in cases that had already been tried, but also in cases that were still ongoing. Thus, because the database is comprised only of cases that were actually judged, it would be reckless to stipulate exactly what would be the size of the reduction obtained, since many cases are still pending trial. Although this may produce a certain artificiality, especially as regards the impact of the plea-bargaining variable on the length of the defendants’ sentences, it is the most conservative approach for the quantitative treatment of these cases.Footnote 8

The first set of independent variables relate to defendant characteristics: their profession, and in the case of politicians, their respective parties at the time. Professions were divided into six groups: politicians, civil servants, business owners, business executives, financial operators and others. The first encompasses both elected officials who did not hold office at the time of the sentence,Footnote 9 as well as high-level party functionaries (e.g., treasurers); the second includes all public sector officials, including executives at state-owned companies like Petrobras; the third includes owners or shareholders of private sector companies; the fourth concerns salaried executives in the private sector; financial operators includes money launderers colloquially known as doleiros and other agents specialized in asset laundering; and finally, the last category is residual and includes unwitting accessories (laranjas), family members and advisors. The political parties of those classified as “politicians” in the previous variable were divided into four groups: PT, PP, PMDB, and others.

The second set of independent variables relates to prosecutorial decisions: the type and number of crimes charged; the amount of money laundered; the types of evidence utilized; the timing of indictments; and the existence of a plea bargain prior to sentencing. Crimes were divided into five types: active corruption (bribe-paying), passive corruption (bribe receipt), criminal organization, money laundering, and others. The latter is a residual category that captures crimes that were charged less frequently (e.g., tax evasion, obstruction of justice, use of false documents, international drug trafficking). The number of crimes for which the defendants were accused is a sum of these dummies (including the category “others”). This generated a continuous variable that ranged between 1 and 6. Types of evidence were categorized as follows: phone intercepts, electronic data, search and seizure, and forensics by the Federal Police (PF) and the Public Ministry (MPF), as well as evidence produced by other agencies, such as the Central Bank (BACEN), Federal Revenue Secretariat (SRF), the Financial Activities Control Council (COAF), the comptroller general’s office (CGU) and the federal accountability agency (TCU), as well as information provided by Petrobras itself and evidence derived from international cooperation. The sum of these dummies accounted for the total number of types of evidence presented in court, ranging from 2 to 10. To account for the accumulation of evidence within the overall operation, the variable “order of indictments” orders the cases according to the date in which indictments were presented. “Number of prosecution witnesses” is a continuous measure ranging from 2 to 29, extracted from Rodrigues (2020). “Plea bargainer” indicates whether the defendant signed a plea bargain before sentencing. “Number of other plea bargainers” sums the number of other defendants within each case who had signed plea bargain agreements, seeking to account for the cumulative impact of plea bargainers on another defendant’s chances of conviction and the severity of sentencing.

Table 1 details the dummy variables and Table 2 details the continuous variables. A few descriptive points are of note. The trial data suggest that Moro was a tough judge In Operation Car Wash, especially by comparison to judges in other corruption cases tried in Brazil. On average, 73.7% of Moro′s decisions were convictions, with average sentences of 113.5 months, or nearly nine and a half years in prison. This conviction rate is much higher than found in other studies of judicial treatment of corruption in Brazil: in the case of sitting politicians judged by the Supreme Court (STF), the rate is below 5% (Falcão et al., 2017; Gomes Neto & Carvalho, 2021); in the case of public servants expelled from the federal administration (predominantly for petty corruption), the conviction rate is around 20% (Alencar & Gico Jr., 2011); in the cases of high-ranking public officials tried within the federal court system, the conviction rate is around 10% (Levcovitz, 2020); and in the case of municipal mayors tried by a specialized panel within state courts, the rate reaches 18% (Bento et al., 2020). Evidence from the judicial decision-making in previous national corruption scandals also stresses the severity of Moro’s rulings: in very few of the most prominent scandals of the 1990s analyzed by Taylor and Buranelli (2007) were the defendants ever convicted.

Table 1 Descriptive statistics, dummy variables
Full size table
Table 2 Descriptive statistics, continuous variables
Full size table

Regarding H1a, although Operation Car Wash is often referred to as an investigation into political corruption, “Politician” was the least frequent profession among the defendants judged by Moro, comprising just over 7% of the total, far behind financial operators, who made up 24% of the cases. The fact that there were few political defendants, of course, should not obscure the fact that they were all convicted in all the cases in which they were tried, reaching a perfect 100% conviction rate. This did not occur with any other variable. One consequence of this perfect conviction record for politicians is that there were no differences in conviction rates between defendants from distinct political parties.

Regarding H1b, there was a clear preponderance of members of the PT among the politicians who were tried. While this has been interpreted as a sign of bias on the part of federal prosecutors, who filed charges against PT politicians more often than against politicians of other parties, it is also the case that the PT defendants heard in the trial court were former occupants of federal elective offices and therefore no longer enjoyed the original jurisdiction of the Supreme Court (STF). This was the case for former President Lula and former ministers Dirceu and Palocci, for example. Many politicians from other parties who were implicated in wrongdoing still held public office during the investigations and their cases were therefore not heard by the trial court, but instead by the STF, which moved slowly on cases of politicians implicated in Car Wash. For example, Eduardo Cunha, former president of the Chamber of Deputies and a member of the PMDB, was only tried by Moro after he was suspended by Congress and he was therefore no longer eligible for the original jurisdiction of the STF. Yet, because the number of political defendants is quite small, it is important to be cautious about drawing conclusions within this group of defendants.

Regarding H2a, although Car Wash is usually associated with corruption, the crime most frequently tried by Moro was actually money laundering, present in 85% of his court’s decisions, followed by active or passive corruption (60%), criminal organization (35%) and other crimes (17%). Because defendants were often charged with more than one crime per indictment (the average is 2 crimes per indictment), money laundering made up 43% of the charges, followed by active and passive corruption (which together comprised 30% of the charges), criminal organization (18%) and others (9%). Scholars have been broadly aware that much of the expertise invested in fighting “corruption” in Brazil since the 1990s has in fact been derived from increasing capacity to enforce money laundering laws, including specialized police and judicial bodies (Da Ros & Taylor, 2022). Lastly, the residual category “Other crimes” has the highest conviction rate, 93.9%, which may be due to the specificity of several of these crimes, for which only one defendant has sometimes been accused in the entire database.

Finally, regarding H2b, evidence provided by the Federal Police and federal prosecutors is the most frequent type of evidence, and evidence from searches and seizures by these two bodies is cited in more than 90% of cases. There is also intense participation by the SRF revenue agency and the Central Bank in cases involving tax and bank data, respectively, accounting for more than 80% of the cases. There is comparatively little evidence provided by the COAF financial intelligence agency, the TCU accounting body and the CGU comptroller general’s office. There is copious evidence provided by Petrobras, which aided the prosecution in several cases, as well as from international bodies, with these two sources providing evidence in more than half of all cases. As a result, various types of evidence have been deployed jointly, averaging more than six types of evidence per case. These descriptive results corroborate something that has frequently been reported in qualitative studies of Car Wash: extensive inter-institutional and international cooperation, unprecedented in previous investigations (Rodrigues, 2020; Da Ros & Taylor, 2022). Finally, plea bargainers constituted a significant group of both indicted and convicted defendants. Indeed, plea bargainers’ trials make up almost a third of the decisions, and other plea bargaining defendants within each case average almost 3 per case, reaching a maximum of 10 in one case. This may be a consequence of the widespread and unprecedented use of plea bargains in Car Wash, made possible by the Law of Criminal Organizations enacted shortly before the beginning of the investigation (Rodrigues, 2020; Prado & Machado, 2022a). It may also result, however, from a relatively small set of plea bargaining defendants who were accused in several cases, such as the doleiro Alberto Youssef, who was tried in 16 different cases, and former Petrobras executive Paulo Roberto Costa, tried in 10 cases. Still, it is noteworthy that even in cases of plea bargaining defendants, the conviction rate does not reach 100%. That is, even though all plea bargainers were convicted at least once, not all cases against plea bargaining defendants resulted in conviction.Footnote 10

4 Multivariate Analysis

Because the dependent variable is measured in two distinct yet interrelated ways, the multivariate analysis relies on a two-part model (2PM). The first measure of the dependent variable is a dummy variable of the decision to convict; the second is a continuous variable of the length of the sentences (in months) imposed on convicted defendants. The second decision is dependent on the first: only convicted defendants receive a sentence. Yet, factors affecting the likelihood of conviction need not be the same ones that impact sentence length. Therefore, our identification strategy requires a selection model that estimates two equations sequentially, given the presence of censored data in the second measure of the dependent variable.

We adopted the 2PM due to its flexibility in comparison to other sample selection models and its fit with the characteristics of our data (Vance & Ritter, 2014). First, it allows all the independent variables used in the first stage equation to also be used in the second stage. That is, it does not require the specification of an “exclusion restriction” in the first stage that need to be omitted in the second stage. Second, the 2PM permits the results in the second stage to be interpreted as actual outcomes, not just as potential outcomes. That follows from the fact that the 2PM treats the censored observations at the second stage as zeros, not as missing cases, given that they comprise non-convictions.Footnote 11

The first stage of the 2PM uses multiple logistic regression to analyze the decisions to convict. Coefficients are reported as odds-ratios. The second stage consists of an ordinary least squares regression calculating the marginal effects of the independent variables on sentence length. In each of the two stages, we ran three different models, resulting in six models in all: models 1A, 1B and 1C are the first stage of the 2PM, reporting the results of the multiple logistic regression; and models 2A, 2B and 2C are the second stage of the 2PM, reporting the marginal effects.

The option to adopt three different models in each stage of the 2PM stems from the challenge that the plea bargaining defendants pose to the models. Plea bargainers are essentially confessed defendants, so we use the sentences these defendants faced before the application of the benefits of their plea agreement, as explained previously. Thus, the first model of each stage (the “A” models) excludes cases of plea bargainers (reducing N to 196 in Model 1, and to 135 in Model 2); the second model (the “B” models) includes all cases (N = 289 in Model 1, and N = 213 in Model 2), but excludes only the variable “plea bargainer”; and the third model (the “C” models) is complete, including both plea bargainers’ cases and the “plea bargainer” variable. In all models, we clustered decisions by repeat defendants and used heteroscedastic-robust standard errors.

In the case of the dummy variables for “Profession”, “Political party” and “Crimes”, the omitted categories are the variables “Other professions”, “Other parties” and “Other crimes,” respectively. As observed above, there is a perfect conviction rate for politicians. Consequently, the independent variable “politician” is a perfect predictor of the dependent variable “conviction,” and thus is automatically excluded from the first stage of the 2PM. In this case, the sum of politicians with “other professions” is adopted as the omitted reference category. Also because of that, the variable “Political party” was excluded from the first stage of the 2PM, as all politicians (from all parties) were convicted. We used the natural logarithm of the amount of money laundered to normalize the distribution of the variable. The variable “Sum of types of evidence” was excluded from the multivariate models because it was highly collinear with the other evidence variables (since it is their sum). Additionally, our multicollinearity analysis found a high correlation of “plea bargainer” with profession: only one politician was a plea bargainer, and few in the category “other professions” were plea bargainers. Said another way: the vast majority of plea bargainers were financial operators, executives and business owners.

The results in Table 3 suggest that some variables strongly determine the chances of conviction. In addition to the perfect “politician” predictor identified in the descriptive analysis, this includes the profession of “financial operator,” which increases the chances of conviction by between 3.4 and 12.6 times, relative to the omitted category. Each additional crime with which the defendant is charged increases the chances of conviction by about 14 times. Given that Moro’s court specialized in money laundering, it is interesting that money laundering led to less likelihood of conviction, other things equal. Regarding the types of evidence, two variables are positively associated with increased likelihood of conviction: electronic data and CGU information. The first increases the chances of conviction between 6.9 and 11.2 times, and the second between 45.1 and 175.3 times. Although the magnitude of the coefficients is very high in the latter case, it is worth remembering it is drawing on a very small number of cases (14), so these conclusions should be read carefully. Surprisingly, the number of other defendants who were plea bargainers decreases the chances of conviction, suggesting that prosecutions may become weaker as more plea bargainers are added. Finally, as expected, plea bargainers increase the chances of their own conviction by 2.2 times, as inferred from the only model in which the variable appears, 1C.

Table 3 Multiple logistic regression, first stage of 2PM
Full size table

The variables mentioned in the paragraph above were statistically significant at the 0.10 level in all models in which they were included. A few other variables were statistically significant in some of the models, but not all. The variable “Active corruption” was significant only in models 1B and 1C, with a negative impact on the defendants’ likelihood of conviction. The variables COAF, TCU and foreign cooperation, in turn, were significant only in one of the models (1A in the first two cases, and 1B in the last case). Whereas COAF is negatively associated to the likelihood of conviction, TCU and foreign cooperation are positively associated to it.

Table 4, reporting the marginal effects derived from the estimation results of both the first-stage logit and the second-stage OLS regression, suggests that some variables have a substantive effect on sentence length, conditional on the incidence of a conviction. Politicians were treated the most harshly, with sentences between 46 and 50 months longer than other professions. They were followed by financial operators, whose sentences were longer than defendants from other professions by 25 to 33 months. The number of crimes charged leads to increased sentences, with each charge increasing the expected sentence by 47 to 50 months. Regarding the types of evidence, electronic data increased expected sentence length by 41 to 45 months. Contrary to initial expectations, evidence from the Revenue Service in cases of lifted tax secrecy led to a reduction of 40 to 44 months. Also contrary to initial expectations, each additional plea bargaining defendant within each case is associated with a reduction of about 6 months. Finally, as expected, plea bargainers saw average prison sentences rise by 23 months in Model 2C. This increased sanction for plea bargainers must be read with caution, though, because it reflects the sentence before any benefits were conceded by the court. In other words, this longer sentence time is artificial, since at the end of the day, plea bargaining defendants actually saw their sentences reduced rather than extended.

Table 4 Marginal effects, second stage of 2PM (only convicted defendants)
Full size table

The variables mentioned in the previous paragraph had statistically significant results in all models in which they were included. There were a few other variables that exhibited statistical significance in only two of the three models tested: accused business owners (significant in models 2A and 2B, as compared to other omitted professions); evidence such as the number of prosecution witnesses (significant in models 2B and 2C), information by Petrobras (significant in models 2B and 2C) and by foreign governments (significant in models 2B and 2C); and defendants from the Workers’ Party (PT, as compared to the omitted category that includes PTB and SDD). Except for the last one, all these variables were positively associated with the length of the penalties applied, with the increase in prison terms ranging from 1.6 to 25 months. In the case of defendants from the Workers’ Party (PT), they had shorter sentences than the omitted category, with average sentences that were about two years shorter. However, we would remind the reader that because there are few politicians in our sample, these findings should be interpreted with care.

Table 5 summarizes the results of our expanded analysis, where we varied the omitted reference categories. Accordingly, we ran 36 additional models in which we adopted as omitted reference categories all possible dummy variables for profession, political party, and crime. That is, in addition to the analysis above where the omitted categories were variables “Other professions”, “Other parties” and “Other crimes”, respectively, we also ran models in which the omitted categories were variables “Politician”,Footnote 12 “Civil servant”, “Business owner”, “Business executive” and “Financial operator” for professions; “PT”, “PP” and “PMDB” for parties; and “Active corruption”, “Passive corruption”, “Criminal organization” and “Money laundering” for crimes. The purpose of this expanded analysis is to check for the robustness of our findings. Due to space constraints, we do not display all these results here. Instead, we include in Table 5 only those variables that were statistically significant in a majority of the models.

Table 5 Variables associated with severity of sentencing (shows variables with statistically significant coefficients in the majority of the models)
Full size table

Once we factor in the results from the expanded analysis, the independent variables that most frequently were significant across all of the analysis (both descriptive and multivariate) were politicians, the number of charges, electronic data evidence, whether the defendants were plea bargainers, and the number of other defendants who were plea bargainers. Except for the last one, all variables were positively associated both to the chances of corruption and sentence length; the last variable was negatively associated to both outcomes. The analysis provides particular support for H1a, H2a and H2b.

With regard to H1a, other things equal, politicians exhibit both higher chances of conviction (all were convicted) and longer sentences compared to all other types of defendants. Financial operators, in turn, exhibit only higher chances of conviction, but not lengthier sentences then other non-politician defendants. Lastly, relative to all possible omitted categories, the professions civil servant, business owner, business executive, and the residual “other variables” do not seem to affect either the chances of conviction or sentence length.

Regarding H1b, because all politicians from all parties were convicted, there are no differences in regards to the chances of conviction. The differences displayed in Table 4 regarding the length of sentences for PT politicians, however, dissipated once we ran the additional models. The negative impact of the variable “PT” was no longer statistically significant once we ran the model adopting as omitted reference categories “PP” and “PMDB’—the “PT” was only significant when the baseline was the residual “other parties” category. This reinforces our conclusion that once politicians were charged in the trial court, there was no significant difference in the severity with which the political defendants were sentenced based on their partisan affiliation.

As for H2a, ceteris paribus, the number of crimes defendants were accused of is consistently associated with an increase in both the chances of conviction and sentence length. Charges of money laundering and active corruption are associated with decreased likelihood of conviction compared to all other categories, but do not affect sentence length. Charges of passive corruption are associated to lengthier sentences, but do not affect the chances of conviction relative to other categories. Lastly, charges of criminal organization and the amount of laundered money did not affect either the chances of conviction or sentence length.

Lastly, regarding H2b, three types of evidence were statistically significant both for the chances of conviction and the length of sentences. Electronic data is consistently associated with an increase in both the likelihood of conviction and sentence length. Plea bargainer is also significant, although it should be read with care, as it was only used in two models (1C and 2C) and is somewhat artificial, as noted earlier, because it is computed before the application of the benefits arising from the plea agreement. The number of other plea bargaining defendants is associated with a decrease in both the chances of conviction and sentence length, suggesting that cases with higher number of plea bargaining defendants may have been interpreted as weaker by Moro. Evidence collected by CGU seems to affect the likelihood of conviction, but not the length of the sentences, although the number of cases is small. Inversely, the number of prosecution witnesses, information from Petrobras and from foreign governments seem to increase sentence length, but not the chance of conviction. Information from the Revenue Service, in turn, seems to reduce the length of the sentences, but not the likelihood of conviction.

5 Discussion and Conclusion

Several conclusions can be drawn from this analysis. The first refers to the severity of Sérgio Moro′s sentences in Operation Car Wash, which led to very high rates of conviction, especially compared to other courts in the Brazilian judiciary. As highlighted previously, the Curitiba trial court stands out both historically and by comparison to its peers. The most significant demonstration of the trial court’s severity relates to its treatment of political defendants: regardless of their party affiliation, politicians in Moro’s court faced a 100% conviction rate, with sentences that were 4 years longer on average than those of other defendants, other things equal. This suggests that the Moro’s court was less concerned than other courts in Brazil, and especially than the STF, with the costs involved in convicting members of the political class. With hindsight, this finding is tragically ironic: the severity with which Moro and Operation Car Wash treated politicians eventually led to blowback and backlash, and politicians worked assiduously both to close down the investigation and to stonewall Moro’s anticorruption initiatives as justice minister (Da Ros & Taylor, 2022).

Secondly, although we tend to think of judicial behavior in corruption cases as a relatively tightly-defined category, the decisions taken by the trial court cover a number of different types of crimes and evidence, with the participation of a significant number of government agencies providing evidence. Corruption charges proper (i.e., active and passive corruption) appear to have been secondary to money laundering, which was the principal crime investigated and prosecuted in Car Wash. The types of evidence brought to bear, in turn, were very diverse, including evidence from the joint Federal Police-prosecutorial task force, but also evidence collected by a significant number of other oversight institutions (COAF, Revenue Service, Central Bank, TCU, CGU), the state-owned Petrobras, and even international partners. This corroborates qualitative conclusions about an increase in coordination within the Brazilian accountability system over the last decade (Taylor & Buranelli, 2007; Power & Taylor, 2011; Carson & Prado, 2016).

Third, this paper suggests that sentencing decisions in corruption cases can be productively analyzed as two relatively autonomous decisions: the decision as to whether or not to convict the accused, on the one hand, and the decision about the length of the sentence imposed on convicted defendants, on the other. Our analysis suggests that the variables that affect conviction do not necessarily affect sentence length in the same manner. In fact, there are a significant number of variables that seem to impact only one or the other decision.

Fourth, despite the importance of taking into account the different factors affecting either the likelihood of conviction or the sentence length, it is important not to miss the more general finding that there are variables associated with both decisions. This is clear in regards to politicians, but also evident in relation to plea bargainers, the number of other defendants who were plea bargainers, and evidence based on electronic data, all of which affect both the chances of conviction and the length of sentences. These last findings are tragically ironic: while electronic data seems to have been fundamental to the severity of the sentences issued by Judge Moro, at least a parcel of the subsequent backlash against Car Wash was made possible because of electronic data. Starting in June 2019, The Intercept published a series of news that used hacked electronic data to show ex parte communications between Moro and the prosecutors, casting doubt on the overall fairness of Operation Car Wash.

Finally, it is essential to highlight the centrality of prosecutors, who are gatekeepers in deciding whether or not to move forward with criminal proceedings. Because legal accountability is a sequential process that has distinct phases such as investigation, prosecution and judgment, future research might look at the factors that affect decision-making in other phases of the process. Judicial decisions are of course highly relevant, but without the decision to prosecute, that is moot. Judicial behavior in cases of corruption, after all, is dependent on the behavior of the members of the prosecution, who decide whom to accuse, and which crimes are charged (Gordon, 2009). While our analysis does not address prosecutorial decisions about whether to prosecute and how to frame that prosecution, this paper has demonstrated that these choices are extremely consequential to judges’ decision-making, deserving greater attention. Variables that have been seen to be very influential here—such as the number of crimes charged, plea bargainer, and the number of other defendants who are plea bargainers, for instance—are all largely dependent on the behavior of prosecutors as they assemble their cases.