CONTEXT: CRIMINAL GOVERNANCE IN THREE ANALYTICAL PERIODS

São Paulo’s experience with the PCC offers a crucial setting for testing how criminal governance can simultaneously reduce violence and improve local economic conditions. The city’s transformation from fragmented, violent criminal competition to hegemonic, rule-based governance presents what Eckstein (Reference Eckstein2000) calls a “crucial case” and Gerring (Reference Gerring2007; Reference Gerring, Boix and Stokes2011) labels a “pathway case” for theory development. The PCC’s emergence and territorial consolidation reshaped both homicide dynamics (Biderman et al. Reference Biderman, de Mello, de Lima and Schneider2019; Feltran Reference Feltran2010; Reference Feltran2012) and the local business environment, providing a rare empirical opportunity to assess how non-state actors affect both domains.

In the late 1990s, São Paulo’s criminal world was marked by highly localized gang competition and extreme violence, often referred to as the “gang wars” period. Violence stemmed largely from competition over drug market rents, where territory changed hands frequently through force. This instability had two compounding effects:

1. 1. High homicide rates: endless cycles of revenge killings and territorial disputes drove homicide levels to among the world’s highest for a city of São Paulo’s size.

2. 2. Hostile business environment: the violence produced severe economic externalities—businesses faced direct threats (e.g., robbery, theft, and closure due to violence) and indirect costs (e.g., reduced consumer mobility, infrastructure damage, and elevated security risks).

Additionally, police forces contributed to instability, often siding with one gang over another based on bribes or personal networks. Meanwhile, prisons were ungoverned spaces, fostering the very criminal rivalries that spilled over into street violence (Biondi Reference Biondi2016; Ruotti Reference Ruotti2016).

By the early 2000s, the PCC had established internal prison hegemony, curbing violence within correctional facilities through rule-based conflict resolution and credible deterrence mechanisms (Biondi Reference Biondi2016). Importantly, this model began diffusing from prisons to peripheral neighborhoods. The “white flag” period marks the start of intentional street-level governance by PCC affiliates, many of whom returned to their home territories after incarceration. Three key governance innovations transformed both the violence landscape and the business environment:

1. 1. Prohibition on unauthorized homicides: violence, particularly lethal violence, became centrally regulated and significantly reduced (Biderman et al. Reference Biderman, de Mello, de Lima and Schneider2019; Feltran Reference Feltran2018).

2. 2. Arms control: the visible circulation of firearms diminished, lowering the probability of opportunistic violence that could disrupt daily life and commerce (Feltran Reference Feltran2018).

3. 3. Drug market regulation: the cartelization of retail drug markets prohibited competitive violence among dealers, standardizing prices and bribes, and creating predictable conditions for informal economic activity (Lessing and Willis Reference Lessing and Willis2019).

4. 4. Consignment-based incentives: by providing drugs on consignment, the PCC aligned dealer incentives with long-term compliance and stability (Lessing and Willis Reference Lessing and Willis2019).

These interventions had dual effects: they drastically curtailed homicides while also reducing transaction risks for local businesses, both licit and illicit.

Starting in 2003, depending on the locality, the PCC established full territorial and institutional hegemony across São Paulo’s favelas and peripheries (Feltran Reference Feltran2008a; Reference Feltran2008b; Reference Feltran2011). The period between 2003 and 2012 represents the apex of rule-based criminal governance in the city. Homicide rates fell sharply citywide, a trend directly linked in time and geography to PCC expansion (Biderman et al. Reference Biderman, de Mello, de Lima and Schneider2019). More subtly, but equally important for this study, the business environment improved in several ways:

• • Reduced violent risk for businesses: retail shops, banks, and small firms faced lower threats of robbery, extortion, and interpersonal violence (Biondi Reference Biondi2014; Dias Reference Dias2008).

• • Stabilization of consumer mobility: residents could safely circulate and engage in market activity, facilitating demand-side economic growth.

• • Protection from police disruptions: by reducing visible crime and using strategic bribery, the PCC minimized police incursions, making neighborhoods more predictable for investment and trade (Blattman et al. Reference Blattman, Duncan, Lessing and Tobón2024; Ruotti Reference Ruotti2016).

• • Expanded informal and formal markets: as trust in neighborhood safety grew, residents and outside actors alike became more willing to invest in businesses and infrastructure, increasing both informal economic activity (proxied by nighttime luminosity) and formal economic indicators such as employment and firm creation.

Scholars have documented a range of PCC governance practices that likely contributed to this dual transformation: theft and robbery prohibitions (Biondi Reference Biondi2014; Dias Reference Dias2008); limits on local boss discretion and intra-gang violence (Feltran Reference Feltran2012); formalized forums for conflict resolution (Biondi Reference Biondi2014); protection for victims of domestic violence (Ruotti Reference Ruotti2016); and the strategic regulation of police–community interactions (Ruotti Reference Ruotti2016). Together, these measures drove down violence and reduced key barriers to economic development.

Figure 1 illustrates the dramatic decline in homicides coinciding with PCC territorial expansion (2003–09). In the empirical analysis that follows, I evaluate whether these governance-induced reductions in violence were accompanied by improvements in formal employment, firm creation, and nighttime luminosity in São Paulo’s favelas.

Figure 1.

The São Paulo Paradox: Homicide Reduction in the City of São Paulo as Criminal Governance Spreads

Note: Number of homicides per year in the city of São Paulo, based on data from Centro de Estudos da Metrópole (2021).

Figure 1. Long description

The Y axis is labeled Homicides and ranges from 0 to 5000 in increments of 1000. The X axis represents years from 2000 to 2020 in 5 year increments. The data line begins at its highest point in 2001 with approximately 4900 homicides. It shows a steep linear decrease until 2005 where it reaches approximately 2300. The downward trend continues at a slightly less aggressive rate until 2008 where it levels off near 1100. Between 2008 and 2011 the line remains relatively flat with a slight dip to 900 in 2011. There is a small peak in 2012 reaching approximately 1250 followed by a steady decline through the rest of the decade. By 2017 the number drops below 1000 and reaches its lowest point of approximately 600 in 2019 before a very slight uptick in 2020.

Notably, there is a contentious discussion in Brazil regarding the role of the PCC in pacifying São Paulo (Freire Reference Freire2018; Justus et al. Reference Justus, de Castro Cerqueira, Kahn and Moreira2018). Some authors, as well as the state security forces, argue that the PCC did not, in fact, decrease homicides in São Paulo, and any improvement observed during this period should be attributed mostly, or solely, to policies implemented by the state (Freire Reference Freire2018). A police prosecutor interviewed for this project emphatically stated that “this whole notion [i.e., the PCC pacified São Paulo] is a conspiracy theory.” In this article, I cannot rule out that improvement in public security practices in the State of São Paulo may have contributed to the reduction in homicides observed in Figure 1. However, the evidence in Biderman et al. (Reference Biderman, de Mello, de Lima and Schneider2019)—showing that the decrease in homicides can be observed at the favela level and is aligned with data on PCC expansion—is arguably stronger and is able to control for any policies being implemented in the state.

Meanwhile, studies such as Justus et al. (Reference Justus, de Castro Cerqueira, Kahn and Moreira2018) and Freire (Reference Freire2018) offer quantitative analyses that are limited from a qualitative perspective. Freire (Reference Freire2018), for instance, inaccurately claims that the PCC only fully consolidated its power in 2006 (see Figure 2) and that the subsequent decline in homicides was uniformly distributed across São Paulo. In reality, the PCC spread in favelas mostly between 2003 and 2006, and homicide rates declined more rapidly and sharply in the city’s favelas—areas that initially had the highest levels of violence (Biderman et al. Reference Biderman, de Mello, de Lima and Schneider2019). Justus et al. (Reference Justus, de Castro Cerqueira, Kahn and Moreira2018), on the other hand, use the number of PCC attacks in 2006 at the municipal level as a proxy for PCC strength, a measure that lacks theoretical justification. They overlook the fact that these attacks were strategically orchestrated and do not necessarily indicate the group’s actual power in those areas. In contrast, Biderman et al. (Reference Biderman, de Mello, de Lima and Schneider2019) find a homicide reduction associated with PCC expansion, employing a more robust identification strategy, leveraging granular datasets, and accounting for state-level policy changes. Their approach is better suited to isolating the impact of the PCC on homicide trends in São Paulo.

Figure 2.

Share of the Areas in the Sample with Reports of PCC Activity

Note: Data on PCC activity come from Disque-Denúncia (Biderman et al. Reference Biderman, de Mello, de Lima and Schneider2019). The data include only areas (Área de Ponderação [APs]) where PCC activity has been reported at least once (purple in Figure 3).

Figure 2. Long description

The line graph plots the share of areas with P C C activity over time. The horizontal X axis represents years from 2000 to 2015 in five-year increments. The vertical Y axis represents the share from 0 to 1 in increments of point 2.

* From 2000 to 2002, the share remains at 0.

* Starting in 2002, there is a steep linear increase, reaching approximately point 35 by 2003 and point 65 by 2005.

* The growth rate slows slightly but remains positive, reaching point 85 by 2006.

* From 2006 to 2011, the line shows a gradual asymptotic curve upward, reaching point 95.

* By 2012, the share reaches 1 and remains at this maximum level through 2013.

THEORY

In this article, I argue that under specific institutional conditions, alternative providers of order and stability—such as criminal organizations—can generate significant economic gains in underserved urban areas. I posit that the institutional mechanisms developed to sustain criminal hegemony and reduce homicides—such as procedural rule enforcement, violence suppression, and market regulation—not only decreased violence but also mitigated the broader negative externalities of criminal disputes. In doing so, these institutions fostered greater predictability and reduced risk, enabling increased trade, investment, and economic development in previously marginalized and violence-prone urban regions.

RESEARCH DESIGN

This study employs a multi-method research design to estimate the causal effects of the PCC’s territorial expansion on economic development outcomes in São Paulo’s underserved urban areas. I implement two complementary empirical strategies to evaluate the PCC’s impact on both formal economic indicators (employment and firm creation) and informal economic activity (proxied by nighttime luminosity). Both approaches leverage the plausibly exogenous, staggered timing of PCC expansion across São Paulo’s favelas and peripheries.

Following the identification strategy developed by Biderman et al. (Reference Biderman, de Mello, de Lima and Schneider2019), I compare units exposed to PCC presence at a given time to geographically and temporally comparable units not-yet-treated—those exposed later to PCC governance. This DiD design exploits staggered treatment timing (Callaway and Sant’Anna Reference Callaway and Sant’Anna2021) to isolate the effect of PCC expansion while accounting for time-invariant unit characteristics and common shocks.

Units of Analysis and Outcome Variables

For formal economic outcomes, the unit of analysis is the AP—a spatial unit aggregating contiguous census tracts with similar infrastructure and sociodemographics (see Figure 3), widely used in Brazilian urban policy (GeoSampa 2016). I focus on 206 APs that contain favelas and experienced PCC presence at some point between 2003 and 2013. I define treatment at the AP level as the first instance where any favela within an AP registers PCC presence. All purple APs in Figure 3 contain at least one favela where I have reports of PCC presence.

Figure 3.

APs in the City of São Paulo Containing Favelas and Reports of PCC Presence

Note: The central green units are much richer areas in the baseline that were never treated. Meanwhile, the staggered sample of interest is made up of APs in poorer areas that have favelas within them and where I observe PCC presence through Disque-Denúncia. I do not directly compare these groups in the regression models as they all focus on the staggered sample that is more comparable. Map is based on Disque-Denúncia (Biderman et al. Reference Biderman, de Mello, de Lima and Schneider2019) and São Paulo City data.

Figure 3. Long description

A geographical map of São Paulo divided into numerous small administrative units. The map uses a two-color legend located at the bottom-left.

* At the center of the city, there is a dense cluster of light green units labeled Never Treated. These units extend slightly toward the South and Northwest.

* Surrounding this central core, the majority of the city is shaded in purple, labeled P C C Observed.

* In the North and Northeast, purple units dominate the landscape with only a few scattered green units.

* To the East, a large, elongated section of the city is almost entirely purple.

* In the South, the map transitions from purple units into several very large light green units at the southernmost tip.

* The West and Northwest peripheries are primarily purple, with a small cluster of green units at the far Northwest edge.

For the informal economic outcomes, I conduct a second analysis at the favela level, using nighttime luminosity (DMSP-OLS) data as a proxy for informal economic development (Zhao et al. Reference Zhao, Zhou, Li, Cao, He, Yu and Li2019). This finer-grained approach leverages the same identification design but focuses on the favela as the treatment unit. Employment and firm data come from RAIS (Ministry of Labor) and are aggregated at the AP level. Nighttime light data are obtained from the National Oceanic and Atmospheric Administration’s Earth Observation Group and processed by the Payne Institute for Public Policy (NOAA N.d.).

Identification Strategy: Difference-in-Differences

The staggered timing of PCC entry across favelas allows me to identify treatment effects by comparing treated units with not-yet-treated units over time (Callaway and Sant’Anna Reference Callaway and Sant’Anna2021; Magaloni, Franco-Vivanco, and Melo Reference Magaloni, Franco-Vivanco and Melo2020). This setup mitigates concerns about confounding time trends. I further improve on prior work (Biderman et al. Reference Biderman, de Mello, de Lima and Schneider2019) by adopting double-robust difference-in-differences (drDiD) estimators, which avoid biases associated with negative weighting inherent in traditional two-way fixed-effects models (Callaway and Sant’Anna Reference Callaway and Sant’Anna2021). This approach estimates the average treatment effect on the treated by cohort and time since treatment, allowing me to construct event-study plots showing dynamic treatment effects.

For model specification, I estimate

(1) $$ Y={\displaystyle \begin{array}{l}{\alpha}_1^{g,t}+{\alpha}_2^{g,t}\cdotp {G}_g+{\alpha}_3^{g,t}\cdotp 1\left[T=t\right]\\ {}+\hskip2px {\beta}^{g,t}\cdotp \left({G}_g\times 1\left[T=t\right]\right)+\lambda \cdotp {X}^{\prime }+{\epsilon}^{g,t},\end{array}} $$

where Y represents the dependent variable of interest, T represents the measurement of time and takes the value of each unit of time (e.g., 2001, 2002, and 2003), and G is the variable that defines the treatment cohort, taking the value of t when the unit was treated for the first time (e.g., if PCC is reported in a favela in 2003, then G = 2003 for that unit). The model assumes that once treatment begins, it cannot be reversed—a condition supported by the data. The estimator calculates the average effect of treatment, e years after implementation, for each cohort treated at that time. The results are grouped and used to plot the event studies that are showcased below (Figures 4– 6).

Figure 4.

Job Creation Subsequent to PCC Entry

Note: This figure presents the event-study specifications described in the model. The results were obtained by the author using Callaway and Sant’Anna (Reference Callaway and Sant’Anna2021) with data from RAIS/Ministry of Labor. The units of analysis are highlighted in Figure 3. The model includes a control for subprefecture, which accounts for location-specific public policies. The estimator does not provide coefficients for controls. The event study displays the evolution of outcomes over time.

Figure 4. Long description

The Y-axis is labeled A T T and ranges from negative 0.5 to 1. The X-axis is labeled Periods to Treatment and ranges from negative 11 to 5. A horizontal dashed line marks the zero point on the Y-axis.

From left to right:

- Pre-treatment phase (periods negative 11 to negative 2): Represented by blue vertical bars with grey circular markers. The markers fluctuate around the zero line, with the lowest point at period negative 10 and the highest pre-treatment point at period negative 8. Most confidence intervals (blue bars) cross the zero line.

- Post-treatment phase (periods 0 to 5): Represented by red vertical bars with grey diamond markers. Starting at period 0, there is a consistent upward linear trend. The point estimate at period 0 is slightly above zero, and each subsequent period shows a higher A T T value, reaching approximately 0.5 by period 5. The confidence intervals (red bars) for periods 1 through 5 are entirely above the zero line, indicating statistical significance.

To validate the parallel trends assumption, I conduct placebo tests and visualize pre-treatment dynamics via event-study plots (Callaway and Sant’Anna Reference Callaway and Sant’Anna2021). Additional tests using the 2000 Brazilian Census (IBGE 2001) data reveal no significant pre-existing differences in social development indicators between early- and late-treated favelas. I also implement tests using OLS and Poisson estimates, and using newer estimators following Borusyak, Jaravel, and Spiess (Reference Borusyak, Jaravel and Spiess2024) to ensure robustness across alternative DiD specifications. Further, I replicate and extend Biderman et al. (Reference Biderman, de Mello, de Lima and Schneider2019) by re-analyzing homicide data with updated methods (results in the Supplementary Material).

ROBUSTNESS TESTS: TIMING OF PCC ENTRY AND OBSERVABLE CHARACTERISTICS USING BRAZILIAN CENSUS DATA

In this section, I explore Brazilian census data to understand whether PCC timing of entry is correlated with pre-existing characteristics of the favelas (measured through the 2000 Brazilian census). I also analyze the 2010 census data to map whether the PCC presence has crowded in state services. I estimate a series of OLS regressions where Y is the outcome of interest measured by the Brazilian census in 2000 or 2010. $ \alpha $ is the constant term, and the various $ \beta $ are the coefficients estimated for each year of PCC entry. The models capture the average difference between favelas that were treated in a given year, compared to the favelas where PCC presence is detected in 2003. $ \epsilon\;\mathrm{is}\ \mathrm{the} $ error term in the equation. The equation and full regression tables are available in the Supplementary Material. As can be seen in Figures 8 and 9, there are no baseline characteristics in the 2000 Census that are systematically correlated with earlier/later PCC entry. This suggests that the PCC did not strategically choose the order of its entry into favelas based on observable characteristics of wealth or labor market conditions. The PCC presence seems to be unrelated to the Social Development variables of interest.

Figure 8.

Can PCC Entry Be Explained by Observable Prevailing Infrastructure Conditions?

Note: This figure presents the OLS estimates (Supplementary Tables A3–A5) obtained from models where the year, when PCC is first observed in a favela, is the independent variable and the Census outcome of interest, measured in 2000, is the dependent variable. Baseline category is 2003. Very low earners refer to the percentage of families earning less than one-half of a minimum wage per capita. Illiteracy is measured as the percentage of those aged 10–14 who are illiterate. High earners are families earning more than 10 minimum wages. The author’s analysis is based on data from Disque-Denúncia, IBGE, and DeInfo.

Figure 8. Long description

A multi-panel plot with two rows and three columns. The Y-axis for all panels is Year of P C C entry, ranging from 2004 to 2012. The X-axis for all panels represents coefficient values from negative 0.2 to 0.2, with a vertical red line at zero. Each data point is a green dot with horizontal error bars.

* Top-left panel: Very low earners - percentage. Points are clustered tightly around the zero line with narrow error bars.

* Top-center panel: Illiteracy - percentage. Points are almost perfectly aligned on the zero line with extremely small error bars.

* Top-right panel: High earners - percentage. Points are centered on zero, with a slight rightward outlier in 2008.

* Bottom-left panel: Access to Water - percentage. Most points are slightly to the right of zero, except for 2011 which shows a significant negative estimate with wide error bars.

* Bottom-center panel: Access to Sewage - percentage. This panel shows the highest variance, with points for 2007, 2009, and 2011 falling to the left of zero and 2012 falling to the right. Error bars are significantly wider than in other panels.

* Bottom-right panel: Garbage Management - percentage. Points are clustered very close to the zero line, showing minimal deviation across all years.

Figure 9.

Can PCC Entry Be Explained by Observable Prevailing Social Conditions?

Note: This figure presents the OLS estimates (Supplementary Tables A3–A5) obtained from models where the year, when PCC is first observed in a favela, is the independent variable and the Census outcome of interest, measured in 2000, is the dependent variable. Baseline category is 2003. The author’s analysis is based on data from Disque-Denúncia, IBGE, and DeInfo.

Figure 9. Long description

A multi-panel coefficient plot with four panels arranged in a two-by-two grid. The shared Y-axis on the left is labeled Year of P C C entry, listing years from 2004 at the top to 2012 at the bottom. The shared X-axis at the bottom ranges from negative 0.05 to 0.1. Each panel features a vertical red line at zero, with green dots representing point estimates and horizontal green lines representing confidence intervals.

* Top-left panel: Socioeconomic Development Index. Estimates fluctuate closely around the zero line. The 2012 estimate shows a slight positive shift with a wider confidence interval.

* Top-right panel: Income level normalized. Most estimates are slightly negative or near zero, except for 2008, which shows a significant positive spike reaching approximately 0.04 with a very wide confidence interval extending beyond 0.1.

* Bottom-left panel: Population density. Estimates are tightly clustered near the zero line with narrow confidence intervals, showing minimal variation across the years.

* Bottom-right panel: Access to bathroom Index. Similar to the income panel, most estimates are near zero, but 2008 shows a distinct positive outlier at approximately 0.04 with a wide confidence interval extending to the right edge of the plot.