Market power and wages: evidence from Brazil

Abstract

Purpose

The paper examines the impact of market power on wages within the context of a developing country, focusing on Brazil.

Design/methodology/approach

With access to matched employer–employee data from Brazil, we first characterized the evolution of the local labor market concentration (Municipality Herfindahl–Hirschman Index [HHI]). Then, we built a fixed-effect model with instrumental variables to verify the association between the local labor market concentration and wages. Finally, a difference-in-difference (DiD) was implemented to verify whether a merger transaction impacted the workers’ earnings in the Brazilian banking sector.

Findings

The paper’s findings suggest that there may be a negative relationship between market power and workers’ earnings.

Originality/value

This research conducted an in-depth investigation of the labor market power in a developing country. As far as we know, our work is the first to evaluate the extension of local concentration in Brazilian formal labor markets and to illustrate its evolution over the last decades. Additionally, when going through the effects of market concentration on wages, we use a new identification strategy that explores changes in the HHI that are caused by national trends in an industry as a source of exogenous variation. Finally, the last part of the paper assesses the effects of antitrust policy on the labor market, a kind of investigation that is still scarce.

Citation

Cavalcanti Gonçalves Ferreira, P. (2024), "Market power and wages: evidence from Brazil", EconomiA, Vol. ahead-of-print No. ahead-of-print. https://doi.org/10.1108/ECON-12-2023-0214

1. Introduction

A growing body of evidence shows that firms’ market power is fundamental to explaining macroeconomic phenomena such as the decline in labor share and the growth of inequality. The predominant approach favors product market analysis. A recent example of this view, with significant impact, is the work of De Loecker, Eeckhout and Unger (2020), which derived a method to estimate markups through the cost minimization assumption. However, although relatively less mainstream, there is also relevant literature that addresses the issue of market power and its impacts on workers through the lens of Oligopsony/Monopsony models.

Markups (firms’ ability to set prices in the product market) and markdowns (firms’ ability to set wages in the labor market) are both components of what is commonly referred to as “Market Power,” as demonstrated Tortarolo and Zarate (2020), Ferreira (2021) and Alpanda and Zubairy (2021). They cannot be disentangled without modeling assumptions, and competition indices on both sides (product and factor markets) can be considered a good proxy for the general market power notion. However, in most countries, particularly in Brazil, a comprehensive approach to market power, including the factor/labor side and its implications for labor market outcomes (wages and inequality), remains largely absent among academics and practitioners. As an exception, we can mention some recent works, such as those of Felix (2021) and Guanziroli (2021), as well as a few articles in the specialized press. Therefore, the main goal of this paper is to help fill in some of the gaps in research about the Brazilian labor market and give evidence that can help guide the country’s antitrust policy.

With an agnostic approach that looks at the labor side concentration as a source or proxy for market power, our work first shows how the Herfindahl–Hirschman Index (HHI) of the local labor market (at the Brazilian municipality level) has changed over time. For this purpose, we use a rich formal labor data source from Brazil, the Relação Anual de Informações Sociais/Annual Social Information Report (RAIS). It is an administrative database with records of each formal employment relationship, matching employees and employers in a very desegregated way (at least at the municipality level). The data allow us to calculate the level of local labor market concentration — the employment-weighted municipality’s average of the HHI from each industry.

The results of this first step show that, in general, there was a decrease in the level of concentration and market power in Brazilian local labor markets between 2000 and 2018. Although a significant part of the territory is still affected by high levels of market concentration (HHI close to 10,000), the distribution of HHI in 2018 had a more uniform profile, with less mass at the extremely high values. Both extensive and intensive changes characterized this reduction in market power. There were decreasing patterns in the distribution mode for the average HHI – due to the increase in the number of establishments in each industry – and for the share of those industries with the highest concentration levels (top 10% HHI).

In the second step, we estimated the impacts of market concentration (employment HHI) on average wages. In traditional monopsony models, market power tends to have a regressive effect on wages. Wage losses from firm market power are caused by either a reduction in the proportion of the marginal product received by workers (wage markdown) or through a lower employment level caused by the markup. These model predictions were empirically tested by regressing the log of average wages (in each labor market, municipality–industry pair) on the respective market concentration indices, control variables and a set of fixed effects.

Identifying “causal” effects of the HHI on wages requires more than controlling for unobserved heterogeneity (by fixed effects). They are equilibrium outcomes. So, there are endogeneity/simultaneity problems between wages and market structure (HHI) (in practical terms, regression errors are correlated with regressors). Therefore, this paper’s main specifications took an instrumental variables approach as an empirical strategy. Our instrument interacts with a concentration in labor markets in the first year of the RAIS series (2006) with the concentration (HHI) growth rate for the sector at the national level, between 2006 and period t. In other words, changes in employment HHI in a specific sector–municipality are instrumented by national variations in the concentration of the same sector.

Our results indicate a robust and economically relevant association between wages, and market structure/market power. In our preferred specification, wage-HHI elasticity is about −0.09. Regressions are robust to different samples and change models’ specifications.

Although we have found a statistical relationship between labor market concentration and wages, interpreting this evidence as causal is still controversial in the Industrial Organization literature. As pointed out by Berry, Gaynor, and Scott Morton (2019), many factors may impact both concentration and market outcomes, and our set of fixed effects and instrumental variables may not rule out all the sources of bias in the regression models. That is why this paper has a third step, where causality between concentration and wages is established through a quasi-experimental research design.

Coping with the tradition of the empirical industrial organization literature, we study a specific market, the banking market, and assess the impact of a merger operation on workers’ wages. Our empirical strategy (differences-in-differences – DiD – approach) allows us to consider the merger of two Brazilian banks as an exogenous shock on affected local markets, and, therefore, obtain treatment effects estimates.

Our results reveal that, after the M&A, the higher market power in the labor and product markets may had a negative effect on mean wages, depending on market size. The merging operation has an overall sensitive impact on the local banking sector level in small markets, with a high concentration level prior to the transaction. Further, in all markets, at the merging firms level, the wage reduction in the treated group ranged from 2% to 3% until four years before the transaction. These values are relatively consistent when controlling for composition effects and heterogeneity among control and treated municipalities.

Our work is related to several others concerned with the repercussions of labor market power. Aside from the previously mentioned papers, it is worth noting Arnold (2019) and Prager and Schmitt (2021), which used DiD approaches in the merger context; Rinz (2020) and Abel, Tenreyro, and Thwaites (2018), which studied local concentration in the US and UK labor markets, respectively; and, finally, Naidu, Posner and Weyl (2018), and Marinescu and Hovenkamp (2019), which discussed how antitrust authorities should deal with anti-competitive labor market conduct. Nevertheless, this work contributes to several antitrust/monopsony literature dimensions.

This research conducted an in-depth investigation of the labor market power in a developing country. As far as we know, our work is the first to evaluate the extension of local concentration in Brazilian formal labor markets and to illustrate its evolution over the last decades. From 2000 to 2018, the Brazilian economy experienced a volatile macroeconomic scenario, with an economic boom between 2006 and 2014 followed by a burst. The market power analysis reveals a markedly anti-cyclical movement during the boom and a moderate one during the crisis.

Additionally, when going through the effects of market concentration on wages, following Rodríguez-Castelán, Lopez-Calva and Cabanillas (2020), we use an identification strategy that explores changes in the HHI that are caused by national trends in an industry as a source of exogenous variation (instrumental variable).

Finally, the last part of this research raises concerns about antitrust policy’s impact on labor markets. Investigations assessing the effects of antitrust policy on the labor market are still scarce. This is our motivation to take a closer look at a merger process released with restrictions by the Brazilian competition authority. To what extent does this operation affect the labor market? Our empirical investigation can provide some answers to improve the merger review process.

The rest of the paper proceeds as follows. Section 2 presents theoretical insights that motivate our research. Next, we show some data aspects and the Brazilian local labor market concentration evolution (Sections 3 and 4). Section 5 presents empirical strategy and estimates of the concentration effect on wages. Section 6 presents the empirical framework, data and findings from our differences-in-differences estimation. Then, Section 7 concludes.

2. Theoretical motivation

We begin this section by recovering De Loecker et al. (2020) derivation of firms’ markup from the cost minimization problem. This approach has been broadly adopted by recent literature about the effects of market power. An optimizing firm chooses its variable input quantity (labor, material) V_i, i = 1, …, N to minimize

(1)C=∑i=1,…,NWi⋅Vi

subject to production technology:

(2)Y¯=FV1,V2,…,VN

W_i is the factor price (wage), V_i is the variable input (labor, in our case), Y is output and F(…) is the production function. λ is the Lagrange multiplier, and so the first-order condition for V_i can be written as:

(3)Wi=λ⋅∂Y∂Vi

Given the Envelope Theorem, the Lagrange multiplier λ should be seen as the marginal cost. Using the fact that marginal cost can be expressed as the ratio between prices and markups (λ=Pu), we can derive a simple formula for the markup:

(4)u=αiSVi

Where SVi is V_i’s factor share of revenue, while α_i is the elasticity of output with respect to V_i. Given its relative simplicity, this ratio estimator has been largely used in the literature (IO, Macro and Trade) to estimate aggregated or sectoral markups.

However, a relevant detail is sometimes ignored. This approach has a fundamental assumption of perfect competition in factor markets. If imperfections give buying power to firms, the minimization problem requires a new formulation, and the markup cannot be disentangled from the factor/wage markdown. As Tortarolo and Zarate (2020) highlighted, the first-order condition of this problem with respect to any variable input is now:

(5)Wi1+1ϵi=λ⋅∂Y∂Vi

Where ϵ_i is the elasticity of factor/labor supply and the term between parentheses is the inverse of the markdown, md_i. So, the initial markup formula now represents a market power index, given by the ratio between markup and markdown:

(6)mkp=umdi=αiSVi

This result shows the importance of factor-side imperfections as a source of market power. In several contexts, it is very imprecise to talk only about markups or product-side market power. Labor market power is also a channel to consider when researchers evaluate market power’s consequences on economic efficiency and distribution. Comprehensively, it is better to refer to the notion of market power, even when using only one side, product or factor markets, estimate or index.

Our empirical work uses the HHI of employment, a measure of labor market concentration, as a proxy for market power to study its impacts on wages. This approach can be theoretically substantiated by several models. We recover the model in Ferreira (2021). It is a dynamic stochastic model with heterogeneous households. For details of its formulation, we refer the reader to the paper. Here, we consider only the supply side and the firms’ problems, which results in an inverse relationship between wages and the number of competitors. Wages levels are given by:

(7)wt=mdtytαlt1μt

Where y_t, α and l_t are output, productivity and labor. With the markup, μ_t(N_t):

(8)μt(Nt)=τ−1Ntτ−ϕτ−1Ntτ−ϕ−1

and wage markdowns md_t(N_t):

(9)mdt(Nt)=τw−1Ntτw−ϕwτw−1Ntτw−ϕw+1

N_t is the number of competitors in the market. Parameters τ_w and ϕ_w are the elasticity of substitution within and across markets, respectively. Both markup and markdown depend on the number of firms. With symmetric firms, is possible to show that HHI=1N×10,000. Since firms’ market power is a combination of the effects of markdowns and markups (as highlighted earlier), we established a direct link between the concentration index (HHI) and wages. The expected behavior of wages can be checked in 1, where we plotted the results from the model’s deterministic steady states, changing only the number of competitors while keeping constant the same parametrization as in Ferreira (2021).

In Figure 1, only the number of competitors changed, keeping other economy’s outputs, in particular the levels of productivity, constant. Further, the model works with homogeneous firms, thereby obtaining pretty clean paths to wages. Nevertheless, in the real world, there is substantial heterogeneity in the size and productivity of firms.

Even considering firm size heterogeneity, the negative impact of HHI on wages remains valid. As shown by Berger, Hasenzagl, Herkenhoff, Mongey, and Posner (2023), with Cournot competition and size heterogeneity, the wage markdowns are a function of firms’ labor share (s_ιjt):

(10)mdtsιjt=sιjt1ϕw+1−sιjt1τw−1sιjt1ϕw+1−sιjt1τw−1+1

However, due to productivity heterogeneity, the observable relationship between concentration and wages may not be straightforward. Since Williamson (1968), it’s been known that there’s a trade-off between concentration and efficiency gains, and from this balance, price movements become ambiguous when there’s market consolidation. The same holds true on the labor side: larger companies may exert greater downward pressure on wages due to market power, yet their higher productivity can lead to rent sharing, a mechanism with an upward effect.

Therefore, the answer regarding the sign of this relationship between market concentration and wages is ultimately empirical and depends on the characteristics of the different times, markets and regions studied.

3. Data

We begin this section with a brief description of the data and some concepts used throughout the paper. First, our preferred definition of the local labor market is the municipality–industry pair. Ideally, we have used the Classificação Nacional de Atividades Econômicas/National classification of economic activities (CNAE) subclass level six digits, at its most recent version (2.0), as our industry definition.

That was not the case with our more extended data series, which was used to describe concentration evolution in the labor market (from 2000 to 2018). The industry classification adopted to harmonize the series was the more aggregated and older version of CNAE (1.0 classes/five digits).

Initially, CNAE classes were the only desegregation level present in the RAIS database. Six-digit codes (subclasses) were made available only from 2006 onward. Additionally, the transition to version 2.0 of CNAE took place in the same year. CNAE 2.0 breaks the previous industries into more sectors; therefore, a different, backward, technique would artificially increase the HHI in recent years, even maintaining the same five-digit level (the greater number of sectors results in “smaller” and, consequently, more concentrated markets).

There is a second relevant issue with the “local labor market” definition. The municipal level may not be comprehensive enough to include all worker mobility patterns. Likewise, the industry may not be the crucial element for substitutability among firms. Most papers with US data use commuting zones (CZ) or core-based statistical areas (CBSA) as a geographic basis (Rinz, 2020; Arnold, 2019). Also, some benchmark papers consider the location-occupation pair as another way of describing a local labor market (Lipsius, 2018; Qiu & Sojourner, 2019; Azar, Marinescu, & Steinbaum, 2020). The concept closest to CZ in Brazil is the microregions defined by the National Statistical Office (Instituto Brasileiro de Geografia e Estatistica (IBGE)) and adopted by Felix (2021) in her work. Additionally, the RAIS database has information about workers’ occupations (CBO codes).

Because of these issues about the definition of the local market, we tested three alternative formulations of the local labor market in the robustness exercises for the wage models: municipality-CNAE class (five digits), microregion-CNAE subclasses (six digits) and municipality-occupation pairs.

As we stated before, our primary data source is the RAIS (Annual Social Information Report) database. It is an administrative register, mandatory for firms and used to pay social benefits for formal workers. There is, therefore, a strong incentive for filling in the information correctly. The database started in the mid-1980s, but we opted for a narrower time frame (from 2000 or 2006) due to changes in the variables over the years. Within RAIS, it is possible to retrieve, through the unique identification, all the worker’s links over the years (hiring and termination date for each firm, monthly and average salary, and occupation), social data such as age, gender and level of education. There are also specific data from firms (at holding or establishment level), such as tax identification number, sector and sub-sector of activity and location (municipality and state), in addition to the number of establishments’ employees.

With data for each establishment (the number of employees, location and industry), we calculated the specific HHI of employment for all “local labor markets.” As usual, the markets’ HHI was obtained by the sum of the squared share (s) of each establishment’s employment, following the equation below:

(11)HHIi,r,t=(s12+s22+s32+…sn2)×10,000

Where i is the sector (CNAE class or subclass), r is the region (in our benchmark specification, the municipality), and n is the number of firms. Therefore, the pair ir is the local labor market. This HHI is the variable of interest on which we regressed average wages in the local labor market.

4. Labor market concentration

The RAIS data also allow us to calculate a kind of local (municipality) index of labor market concentration: the employment-weighted municipality’s average of the sectoral HHI. The weighted and aggregated municipality HHI version reads:

(12)HHIr,t=∑iWeighti,r,t×HHIi,r,t

The Weight_i,r,t is calculated by the ratio between the local industry and total municipality employment. From this local HHI, we could also obtain the “national” index of local market concentration by aggregating once again, this time by population weights.

(13)HHIt=∑rPopulationr,tNat.Populationt×HHIr,t

The last aggregation results are plotted in Figure 2. The panel shows the evolution of local employment concentration from 2000 to 2018. Between the second half of the 2000s and the first of the 2010s, the Brazilian economy was hit by a positive shock caused, among other reasons, by the commodities boom. However, the economy slowed down in 2014, and, from 2015 onward, the country faced a severe recession. In the literature, there are authors, like Lambrecht (2004), who point to a positive correlation between economic expansion and merger and acquisition activity, which would ultimately lead to an increase in market concentration during economic booms. Instead, considering the stylized facts in 3, the concentration trend in the Brazilian local labor market seems to correspond more to the behavior predicted in macroeconomic models outlined by Jaimovich and Floetotto (2008), Etro and Colciago (2010), and Ferreira (2021), which predict an inverse relationship between growth and market concentration. Until 2015, the concentration trend was firmly downward during the Brazilian economic boom. In the following years, there was a reversal, although the upward movement was moderate.

The territorial pattern of labor market concentration and the municipal HHI density can be seen in Figure 3 and in the left panel 4, respectively. The two ways of viewing the same data reinforce the finding that the labor market in Brazilian municipalities has become less concentrated in recent decades. At the end of the series, in 2018, we see fewer red spots on the Brazilian municipalities map, indicating that several cities moved from a very high level of market concentration (HHI between 7500 and 10000) to a high (between 5000 and 7500) or moderate concentration (2500 and 5000). In other words, when compared to 2000, the 2018 municipal HHI has a much more uniform profile in the middle to the right of the distribution in Figure 4, with less mass at extreme values.

Despite deconcentration in the labor market, it is still possible to find regions, especially in Brazil’s North and Northeast, where the HHI reaches very high values, especially in smaller locations. Some are conceptually considered in a perfect monopsony condition (HHI 10000), where the only formal employer is the municipal administration.

Our municipality’s concentration index was obtained from the sum of local industries’ HHI, weighted by their shares of total employment. So, changes in the concentration distribution can be caused by extensive variation within industries, i.e. a reduction or elevation in the economy/municipality’s mean HHI, or by between variation, with increasing or decreasing patterns in the shares of the most concentrated industries.

Inspecting Figure 4, it is worth noticing that the deconcentration trend in local labor markets is due to both types of variation. The panels on the right represent, respectively, the distribution of the average HHI from municipalities’ industries and the employment share of those 10% with the highest level of concentration. From 2000 to 2018, we observed a reduction in modal values from both distributions, with tails on the right with less weight. Going further, it is possible to investigate why there was a general reduction in the average HHI of local economic sectors. This question is addressed in Figure 5.

The sectoral HHI at the municipality level is calculated as the sum of the employment shares of each firm. Therefore, it is influenced by the number of employees in the establishment and the number of competitors operating in a given sector and location. The left panel shows the distribution of the average number of firms in local industries (weighted by sector importance in total municipal employment). The second panel, on the right, shows the distribution of the average number of employees per firm among Brazilian cities.

Brazil experienced accelerated growth in the period under analysis. Consequently, there was an intensive labor market expansion (increasing the average employment level per firm). On the other hand, the same period was marked by a significant entry movement, illustrated by the rise in the modal values from the number of firms’ distribution. If there wasn’t such extensive growth in entry, which offset the increase in incumbent firms’ employment, there would be a greater market concentration due to the economic boom.

Also, between 2000 and 2018, the country witnessed a cycle of advances and setbacks concerning workers’ wage levels. How can these phenomena be partially explained by the market concentration effect? This evidence about this question will be revealed in the next session, in which we outline an empirical strategy for estimating the impacts of HHI variation on average sectoral wages.

5. Concentration and wages

5.1 Model and IV

Our model is a reduced-form fixed-effect regression of wage averages on market HHI (and a vector of control variables) at the local labor market (industry-municipality level), with the following specification:

(14)Log(Wage)i,r,t=β1Log(HHI)i,r,t+∑qβqXq,i,r,t+LaborMktir+Yeart+εi,r,t

Its components are indexed by i for industry, r for region (municipality) and t for time. Log(Wage)_i,r,t is the log of the local labor market (industry-municipality) average wage in a given year. Log(HHI)_i,r,t is, consequently, the industry-municipality employment concentration index (and our coefficient of interest is also β₁, the HHI elasticity of sector mean wage), while ∑^qβ_qX_q,i,r,t is a vector of control variables (average values for tenure, gender proportion, education, age; and municipality population). Finally, there is now a set of fixed effects controlling for unobserved heterogeneity: LaborMarket_ir (Region_r interacted with Industry_i) and Year_t.

Employment concentration levels were obtained from RAIS, computing the sectoral HHI index. However, the use of data from 2006 onward allowed us to consider a narrower industry classification level, the CNAE’s subclasses (six digits). There is no census information for each labor market, so we rely on RAIS to obtain the remaining variables (average wages and controls).

The data spans from 2006 to 2018, forming a panel with 13 time points (12 in the instrumental variable model). In total, our complete sample has 6,919,451 observations (532,265 labor markets per year, on average), resulting from yearly data combining about 1,300 sectors and 5,500 locations (CNAE subclasses codes related to public administration were excluded from our sample). Descriptive statistics for all data series are reported in Appendix A (Table 6).

Not all municipalities have observations for all industries (the mean number of industries per municipality is 96.79). Additionally, not all local labor markets (industry-municipality) are observed in the 13 years of the series, giving us an unbalanced panel. To verify if this issue impacted our estimation, we will test some specifications with only intensive changes in the sample (i.e. only markets that appear in all years).

The validity of fixed effects estimates assumes that omitted heterogeneity is the only source of bias. This assumption is probably wrong in our current model because the endogeneity concerns (originated by double causation and omitted variables, given that wages and HHI are market equilibrium outcomes) are relevant. So, our preferred estimation procedure adopts a quasi-exogenous source of variation to instrument the local employment HHI. As in Rodríguez-Castelán et al. (2020), our analysis relies on an instrumental variable strategy. The local industry HHI for employment in a municipality is instrumented by national changes in the concentration of the same sector. That is, the sector i and municipality r specific HHI instrument is defined as below:

(15)IVHHIi,r,t=HHIi,r,t=2006×gi,t

Therefore, our instrument is computed through the interaction between two pieces of data: HHI_i,r,t = ₂₀₀₆, which is the concentration in labor market ir (sector–municipality pair) in the first year of the RAIS series (2006); and g_i,t, the concentration (HHI) growth rate for industry i at national level between 2006 and period t. Using fixed local concentration levels (HHI_i,r,t = ₂₀₀₆), which we can assume as exogenous to the forthcoming years, the instrument rules out specific and unobserved local shocks affecting concentration levels and wages simultaneously (the source of bias). Moreover, the instrument brings back some variation with the supposedly exogenous changes in the local labor market concentration (g_i,t) driven by HHI national trends (caused by national trade policies, industry-specific incentives, the competitive environment or the labor market). First stage F tests support the instrument’s relevance in predicting local employment HHI (the complete first stage results are in Appendix, Table 8).

Our benchmark estimation originates from a non-outlier sample – removing local markets with zero, top, or bottom 1% mean earnings. This procedure reduces concerns about measurement errors. One of the model’s robustness tests measured the impact of this choice.

5.2 Results

The main results are shown in Table 1 (complete regressions results are in Appendix, Table 7). Overall, as predicted by several models, there is a negative and statistically significant relationship between concentration and workers’ earnings (except for the non-instrumented TWFE regression, which does not reject the null). The result holds and becomes more substantial when estimation relies on IV strategy. In this specification, the preferred one, a reduction of 10% in local labor market concentration leads to an economically and statistically significant increase of approximately 0.8% in the mean wage. Moreover, the instrumental variables approach indicates that the model which does not control for unobserved local shocks is biasing toward zero the true effect of local labor market concentration on earnings. In fact, in the non-instrumented specification with a time fixed-effect, the coefficient is insignificant and zero to the third decimal place.

The positive bias deserves to be explored further. It may be caused by measurement error. However, recovering the model that theoretically underlies this work, it can be speculated that local demand/productivity shocks may play a role, affecting more concentrated sectors, and increasing both the HHI and the wages (through marginal revenue). Unfortunately, there is no firm-level data on production and revenue in the RAIS database. The Brazilian statistical office (IBGE) has this kind of data in its annual sectoral censuses. However, the identified base is confidential, and we could not access it timely [1]. Despite not quantifying the impact of firm productivity on the wage model, it is possible to have confidence in the evidence obtained as the main goal of the instrumental variable strategy is to deal with the bias from unobserved local shocks.

The last three columns of Table 1 present robustness test models in which we changed the definition of the local labor market. The first one documents the estimates in which the HHI was calculated at the microregion level (i.e. the labor market considered is the industry-microregion pair). It attempts to use a geographical component closer to what is commonly used in papers with data from the United States, the CZ. The second estimates a model with a more aggregated industry classification (CNAE five digits or classes). The last considers no longer the firms’ industry but the occupation of workers as a reference. Due to the computational burden, this occupation model relies only on a data series from 2015 to 2018.

There are statistically significant numerical differences in both models concerning baseline estimates. Wages are considerably more elastic when the market concentration index considers occupations. In the opposite direction, the estimate for microregions and CNAE class are lower. These variations are in line with expectations as they are computed in smaller and larger markets, respectively. Nevertheless, it is noteworthy that baseline and alternative estimates are qualitatively similar and reaffirm the evidence that market power dampens Brazilian workers’ earnings.

In addition to changes in the local labor market concept, our estimates were subjected to four other robustness checks. They are reported in Table 2. Columns (1) to (3) show the instrumental variable’s results on three different samples. The first is the complete sample, recovering outlier observations. In the second, observations from monopsonistic sectors (HHI = 10000) were excluded. Finally, considering that not all local labor markets (industry-municipality) are observed in the 13 years of the series (an unbalanced panel), the third sample consider only those markets which have observations for those labor markets with at least one firm in all years of the sample’ time span.

We changed the regressions specifications in models (4) to (6). (4) has an auto-regressive term (Log(Wage)_i,r,t−1), while in (5) and (6) we added year-municipality fixed-effects and industries-year average wages (we did not include year-industry fixed-effects (FEs) because as our instrumental variable (IV) relies on national HHI trends they are collinear). Qualitatively, there are no remarkable changes in relation to the preferred regression model.

Finally, in Table 1, there’s a noticeable divergence in the number of observations between the one-way fixed-effect(OW)-two-way fixed-effect (TWFE) and IV specifications. This divergence stems from how the instrumental variable is defined (we can’t calculate the instrument for markets not observed at the series outset in 2006). To assess whether the discrepancy in estimations is methodological rather than due to the sample, we ran the OW and TWFE models using exclusively observations from the IV sample. The outcomes, detailed in the Appendix (Table 9), show no qualitative departure from the initial findings.

Our empirical strategy found statistically robust evidence about the relationship between market power and wages. Nevertheless, before making causal statements, we should consider an open debate about the HHI and its validity as a market power proxy. The HHI is still widely used today, particularly by antitrust authorities, as one of the main ways of measuring market power. The usefulness of the index lies in its simplicity, requiring data only on sales/employment for all firms in a given market. Moreover, there are reasonable theoretical arguments that justify its use, in addition to the model presented in our theoretical motivation (Miller & Sheu, 2021), and it appears to have good predictive power regarding market outcomes, as we see in Autor, Dorn, Katz, Patterson, and Van Reenen (2020). Despite this, there are relevant arguments against the direct relationship between the HHI and the firms’ market power. A summary of them can be found in Eeckhout (2021). The author asserts that the measured HHI crucially depends on how we define a market and that there is no obvious way to do it. Moreover, like other researchers in Industrial Organizations, Eeckhout highlights that concentration is an endogenous outcome with no straightforward instrumental approach.

Aware of the possible limitations of the methodology used so far, the next section will forego the cross-market aggregate analysis, focusing on a specific market, the Brazilian banking market, to assess the effects of a merger and acquisition operation. Adopting a quasi-experimental approach (DiD), it is possible, with some assumptions, to treat the merger as an exogenous shock on affected local markets and, therefore, obtain estimates for the impact of growing market power/concentration on workers’ earnings.

6. Concentration and wages in the banking sector

Our proposed empirical approach involves estimating the treatment effect of the merger between two Brazilian banks. Once again, we relied on restricted data from RAIS, therefore firms’ identities and some other details will not be disclosed. The transaction involved firms that held, on average, close to 25% of market employment. In some smaller cities, this share was up to 75%.

The merger would result in an expected increase of about 200 points in the national employment’s HHI (ΔHHI = 2 × s₁ × s₂). According to the data from RAIS, this index was close to 1,800 in the year prior merger. Taking as reference the guide for horizontal merger review (CADE, 2016) from the Brazilian antitrust authority (Conselho Administrativo de Defesa Econômica/Administrative Council for Economic Defense (CADE)), through the lens of the buying power, the banking labor market could be considered moderately concentrated (HHI ranging from 1,500 to 2,500 points).

On local labor markets, the impact on concentration was quite significant – we found a median ΔHHI of 474 after the merger, nearly five times the baseline of CADE’s horizontal merger guide (100 points).

6.1 Empirical strategy

To estimate the treatment effect we implemented a dynamic DiD technique (saturated TWFE DiD). This approach takes a set of assumptions (parallel trends and others) about the treated group and its counterfactual (untreated) to obtain a natural quasi-experiment, i.e. a supposed exogenous shock that allows evaluating the effects of the increase in market power in the affected municipalities.

In the empirical implementation, the crucial point is identifying the counterfactual group (untreated), which allows estimating the effect on those treated (averege treatment effect on treated(ATT)). In our setting, the transaction occurred at the national level, not being affected by purely local market factors. Furthermore, the different local labor markets (the banking sector, six-digit CNAE code, in each municipality) are competitively heterogeneous: there are cities in which the two merging firms overlap; in others, only one operates. So, these conditions create two municipality groups: the one exogenously affected by the merger (treated), regions where the firms operate simultaneously and the one where initial competitive conditions remain unchanged (untreated), i.e. only one firm had establishments prior to the merger.

The evolution of workers’ mean wage (in both merging firms) for treated and untreated groups can be inspected in Figure 6 (where the vertical dotted line indicates the year prior to the merger announcement), and some descriptive statistics about them are in Appendix (Table 10). It is worth noticing that, before the merger announcement (until time −1), there was no evidence that the two groups were trending differently, although there were some non-negligible differences between the municipalities (the treated cities are, in general, larger markets with a bigger GDP; they also have older workers, with a higher percentage of women; however, median wages vary little between treated and untreated).

6.2 DiD models

We estimate DiD for two different outcomes: the market and the merging firms’ log of mean wages (both, in each year, at the municipality-establishment level (Log(Wage)_ri,t). For each outcome, three variations of the general differences-in-differences model were implemented. The first one is the simple TWFE, in which there is only one parameter of interest (δ), linked to the interaction between the dummy of the treatment group and the post-treatment period. The model’s specification can be checked in equation (16).

(16)Log(Wage)ri,t=δTreatr×Postt+Regionr+Yeart+εri,t

As usual, Region_r and Year_t are the terms that control for municipalities (treatment level) and time-fixed effects. There is extensive literature regarding the limitations of the simple TWFE approach. The problems identified by multiple authors (Goodman-Bacon, 2021; Callaway & Sant’Anna, 2021) are, in particular, related to settings where there are multiple periods (dynamic effect) and variations in treatment timing (staggered adoption). Concerns about different treatment timing are not an issue in our empirical implementation. However, the merger-effect dynamics (multiple periods) are relevant to us.

Wages, prices and market shares can be more or less flexible, and the variation in these outcomes’ rigidity determines the evolution of wages in our sample periods. We chose to set the period after the merger announcement (not its approval by the antitrust agency) as the treatment “beginning,” because, while there are regulations against what is known as gun-jumping, pre-merger coordination between the parties can occur, for example, concerning pricing, without the authority realizing it. On the other hand, hidden unlawful pre-approval changes in employment contracts would be rare, as unions monitor them.

This dynamic may have an identifiable impact on wages: if prices go up, marginal revenue goes up, and, if there is some rent-sharing process, wages can be higher in the treated group, despite the market power; later, with firms’ monopsony power, wages can be pressured downward. The simple TWFE estimator would hide this post-merger wage behavior. After the estimation, we get an aggregate value for the multiple periods, as the model concentrates the whole treatment effect and its dynamics on a single parameter. In cases like the one described above, the result would be misleading.

Wooldridge (2021) makes a similar point. For the author, the TWFE model is not necessarily wrong as long as it allows for greater heterogeneity in the treatment effect. This heterogeneity is captured in a saturate specification, including terms for post-treatment periods (dynamic effect). This extended version of the TWFE also accommodates time-invariant or pre-treatment covariates (demeaned by treatment status groups’ mean and interacted with the treatment dummy) and linear/non-linear trends. This approach was adopted in the second difference-in-differences specification, presented in Equation (17).

(17)Log(Wage)ri,t=∑k=04δk1[t=k]×treatr+Regionr+Yeart+εri,t

The new parameters (∑k=04δk1[t=k]×treatr) can identify the treatment effect dynamics.

Furthermore, Table 10 (Appendix) shows non-negligible differences between the two studied groups (treated and untreated municipalities). Because of this, and to make sure that there were no pre-trends, we estimated two other extended TWFE with linear trend terms and covariates, respectively.

Finally, we also implemented an event study style regression [2]. It is an extended TWFE model, but with leads, i.e. the interaction between the treatment’s dummy variable and indicators for the years before the merger (q ∈ {−3, −2}). The specification is detailed in equation (18). The event study approach has two purposes. First, it allows another way to test the pre-trend. Through a joint significance test (Wald test), it is possible to verify the null hypothesis that the coefficients of the leads are equal to zero. Second, it is a good visualization tool for the outcome dynamics.

(18)Log(Wage)ri,t=∑q=−3−2δq1[t=q]×treatr⏟Leads+∑k=04δk1[t=k]×treatr⏟Lags+Regionr+Yeart+εri,t

6.3 Results

6.3.1 Market effects

Similarly to the previous exercise, where we analyzed the effects of market concentration on a local labor market, our first set of results presents the outcomes of DiD regressions on average wages in the banking sector across all affected municipalities. The results are displayed in Table 3 (simple and extended TWFE) and Figure 7 (event study).

The aggregate treatment effect estimated in the simple TWFE model, although positive, is very close to zero and therefore not statistically significant. At first glance, this would seem to show that the merger had no effect on the wages of the workers in the banking market. Nevertheless, it is misleading, as predicted above.

The scenario is richer in the extended TWFE models, which break down the treatment parameters to capture the outcome dynamics. In specification without controls and linear trend, between the merger announcement and the antitrust authority’s approval (from period 0 to 1), there is a positive jump in the mean wages in markets affected by the merger. This positive effect vanishes immediately after the transaction’s clearance when workers’ earnings fall 1.4%. However, these downward movements may be result from previous trends, as we can see in the model (3). When we include a linear trend, the banking labor markets seem to return to the initial state after the CADE’s approval — the model indicates null effects after the initial rise.

However, the concentration effects may be heterogeneous (non-linear) between labor markets. Monopsony models, as in Ferreira (2021), predict that the impact is more pronounced in markets with prior higher HHI (lower number of competitors). That’s why we ran the same extended TWFE model (with covariates and linear trends) only in smaller and more concentrated markets (small regions model, with less than 25 establishments). In these markets, the merger has generated lower aggregate wages, with considerable reductions of around 7% in the last sample period.

The evidence that the merger has more negative repercussions in smaller regions with concentrated labor markets is very much in line with the findings of Prager and Schmitt (2021). Their study showed that the effects on wages were only significant in markets with previously high HHI levels.

In our study, at the national level, the two involved banks held 20% (acquirer) and 5% (acquired) of the banking labor force. In theory, and as evidenced by the data, the impact on the market should be small, in general. However, when focusing on specific cities, where competition was already lower, it was noticed that the two firms held up to 75% of the workforce. The effective HHI variation after the merger had a median of 195 at larger treated municipalities, while the smaller one had a median HHI change of 620. Our findings highlight the need for regionalized merger analysis when focusing on the labor market.

6.3.2 Merging firms

Depending on the manner in which strategic interaction among companies unfolds, variations in prices within a consolidating market may diverge, potentially impacting those closely intertwined with the operation to a greater extent. Consequently, it is customary for antitrust authorities to conduct distinct analyses regarding price implications: on the market as a whole and on the involved companies. There is no doubt that the same procedure holds relevance for the labor market. Hence, we adopt the same DiD methodology to scrutinize the effects on wages paid by the companies involved in the merger, which may have been mitigated in the overall market analysis.

Simple and extended TWFE results (including linear trend and covariate models) are described in Table 4. The estimates from the event study (obtained with and without covariates), in turn, are plotted in Figure 8.

As before, between the merger announcement and the antitrust authority’s approval, average wages went up in establishments affected by the merger. This positive effect vanishes immediately after the transaction’s clearance when workers’ earnings fall between 2% and 3% in regions where market concentration increases (and the effect is persistent until at least period 4).

This wage behavior justifies our choice to consider the merger announcement as the treatment beginning. There is an evident anticipation movement before the approval. How to interpret these up-and-down phenomena? Our view is that retail market prices reacted immediately to the merger expectation, while adjustments in quantities (of products and production factors) may have been delayed. The temporary increase in workers’ pay could be justified by the extra profit the market misalignment brought in (especially in the banking sector, where a big part of pay is tied to performance).

Still, in our initial results, in Table 4, there is no evidence of divergent trends between treated and untreated groups before the merger. The linear trend coefficient is close to zero and statistically insignificant (third column). The Wald test in the event study plot (Figure 8) points to the same conclusion, with p-values higher than 0.20. The covariates also do not have an apparent influence on estimates. The parallel trend assumption seems to hold unconditionally.

In addition to the baseline TWFE models (simple, with and without covariates and linear trend), estimated in the entire sample of municipalities, new estimates (from the model without covariates and trend) were produced in a new sample, composed only of municipalities with less than 50 workers in merging firms’ establishments (small regions model).

The small-region model reproduces the same wage dynamic pattern, with an initial jump followed by a significant drop after the merger’s approval, although the effects are more substantial. It is evidence that the baseline estimates do not violate the common support principle. It is also indicative that there exists a non-linear effect of increasing market concentration. The impact is greater in smaller and more concentrated markets (achieves a 5% reduction in workers’ wages).

Our baseline estimates rely on municipality-establishment’s average wages outcomes for both firms. When running our merging firms’ DiD models, we capture two distinct sources of variation: changes in individuals’ earnings and changes in the composition of the establishments’ workforce. Therefore, the aggregated outcome is subject to composition effects. As Guanziroli (2021) has shown, the merger and its consequent impact on firms’ objective functions would change the proportion of workers from different occupations with different wage levels.

Nevertheless, contrary to Guanziroli (2021), this fact does not lead us to adopt a regression with individual worker data. Our choice is justified, among other reasons, because, when using individual data, we would face a series of other sources of bias, such as, for example, the attrition caused by the merger (endogenous selection). In addition, throughout the entire sample period, some individuals enter and leave the establishments affected by the merger. We would then have an experiment design with variation in treatment timing, but once a unit becomes treated, there is no guarantee that it will remain treated in the next period. In this kind of setup, estimation needs a new set of assumptions that aren’t met by our TWFE model.

Additionally, and perhaps more relevant for the paper’s proposal, variation in aggregate earnings due to changes in the occupation’s proportion is mainly driven by the merger impact. In the end, firms change their optimizing outcomes in response to new levels of buying and product-side power generated by the transaction. Therefore, composition changes caused by market power should not be excluded when evaluating the overall merger effect. However, our benchmark estimation may still be biased by other non-market power composition variations, which are mainly caused by post-merger restructuring. Merging firms may eliminate the redundant workforce, impacting treated establishments/municipality’ mean wage.

This issue was addressed by a new set of estimates using “stayers” workers; i.e. mean outcomes were calculated using wages from individuals who remained in their initial establishments during all the sample periods. With this procedure, we eliminate not only the composition effect caused by the firms’ restructuring but also the merger-induced one. Therefore, our results should be read as a partial identification approach. Stayers’ sample estimates are the lower bound for the treatment effect. At the same time, the benchmark, or entire sample, results are the higher bound.

Stayers’ extended TWFE results (baseline, linear trend, covariate and small regions models) are described in Table 5. The event study’s estimates (with and without covariates, small regions) are plotted in Figure 9. In the models with all municipalities, we cannot exclude the possibility of divergent trends for treated and untreated groups, even in the presence of covariates. Thus, it is recommended to consider the results in the specification with a linear trend parameter (second and third columns). They point to the same initial wage jump, followed by an effect close to zero (not statistically significant). In the small-region model, which controls for violations of the common support assumption and captures the merger impact in previously more concentrated regions, the merger effect estimates remain strongly negative, even in the presence of the linear trend. When the results from the initial benchmark and the stayers’ model are added together, they show that the merger had, at best, no effect on workers’ wages. However, in all outcomes (market and merging firms average wages) there is some evidence that workers’ earnings went down, at least in smaller and more concentrated markets.

7. Conclusion

The effects of market power on wages remain absent from the literature and antitrust practice in developing nations, particularly Brazil. The objective of this study was to provide evidence to guide potential modifications to the way antitrust authorities approach labor market issues.

With detailed and identified matched employer–employee data from Brazil, it was initially possible to characterize the temporal evolution of the local labor market concentration (municipality HHI, an index constructed with the weighted aggregation of the local industries’ HHIs). Then, we constructed a fixed-effect model with instrumental variables to examine the relationship between local labor market concentration and wages. To address possible concerns about the validity of the regression models with the HHI as the independent variable, a quasi-experimental empirical strategy was used to determine if the increase in market power caused by a merger and acquisition operation in Brazilian banking had an effect on workers’ wages.

Our findings support the policy recommendation that Brazilian antitrust authorities, concerned with the impact on workers’ earnings, should increasingly incorporate labor market data into their analysis procedures. Despite being included in CADE’s manuals, this type of study is uncommon in Brazilian M&A reviews. Equally uncommon are actions against employment contracts with non-compete clauses or no-poaching agreements between competing firms. Reforming antitrust processes creates a vast area for future research that aims to aid practitioners by focusing on the development of new policies, the creation of labor market tools and models, and the formulation of antitrust remedies. Several articles on this subject have been published, including those by Rose (2019), Sillman (2020), Naidu et al. (2018) and Berger et al. (2023).

There is also some empirical and academic work to be done on the relationship between market power and wages. This literature continues to rely heavily on evidence derived from reduced-form regressions. There is room for works with structural specifications, such as Felix (2021), to estimate a more comprehensive set of parameters and interrelationships between firms’ and workers’ decisions. This domain is also compatible with the causal inference framework utilized in our research. More post-merger or post-cartel analyses of the responses of wages in various markets will contribute significantly to policy guidance. Finally, we studied market power without effectively separating factor (labor) and product-side wage effects (although using employment HHI as our market power proxy). For the Brazilian case, research papers could be written that disentangle the effects of each market power’s component (markups and markdowns). It will need access to identified databases on the labor market, such as RAIS, and to sectoral censuses conducted by the Brazilian statistical office (which contain information on firms’ production and revenues).

Figures

Figure 1

Model simulation – number of firms and wages

Figure 2

HHI evolution – Brazilian municipalities, 2000–2018

Figure 3

Map – HHI evolution – Brazilian municipalities, 2000–2018

Figure 4

Densities-HHI evolution – Brazilian municipalities, 2000–2018

Figure 5

Number of establishments and employees – Brazilian municipalities, 2000–2018

Figure 6

Treated and control: mean wage evolution

Figure 7

Event study estimates – market

Figure 8

Event study estimates – merging firms

Figure 9

Event study estimates – merging firms and incumbent workers

Table 1

Effect of market concentration on wages (mean by CNAE subclass)

Dependent variable	log(Wage)
	OW FE	TW FE	CNAE subclass	Micro	CNAE class	Occup
Model	(1)	(2)	(3)	(4)	(5)	(6)
log(HHI)	−0.054***	0.000	−0.082***	−0.060***	−0.047***	−0.121***
	(0.0005)	(0.000)	(0.005)	(0.004)	(0.005)	(0.018)
Mun-Ind (subclass) FE	Yes	Yes	Yes	–	–	–
Year FE	–	Yes	Yes	Yes	Yes	Yes
Micro-Ind (subclass) FE	–	–	–	Yes	–	–
Mun-Ind (class) FE	–	–	–	–	Yes	–
Mun-Ocup FE	–	–	–	–	–	Yes
IV	–	–	Yes	Yes	Yes	Yes
Observations	6,628,761	6,628,761	3,854,965	1,739,115	3,208,241	3,355,276
Kleibergen-Paap Wald test (1st stage)			10904.4	9509.4	8436.2	1072.3

Note(s): Signif. Codes: ^***: 0.01, ^**: 0.05, ^*: 0.1

Standard-Errors clustered by labor market (id)

Source(s): Table by author

Table 2

Robustness – effect of HHI on wages

Dependent variable	log(Wage)
Model	(1)	(2)	(3)	(4)	(5)	(6)
log(HHI)	−0.090***	−0.069***	−0.091***	−0.066***	−0.090***	−0.105***
	(0.005)	(0.005)	(0.005)	(0.005)	(0.005)	(0.005)
Sample	full	non-monopsony	balanced	benchmark	benchmark	benchmark
Add. vars	–	–	–	AR1	Mun-Year FE	Avg. Ind. wage
Observations	4,011,447	2,587,021	2,639,076	3,497,292	3,854,917	3,854,917
Kleibergen-Paap Wald test (1st stage)	11555.4	7784.7	7300.1	8062.6	10684.1	10627.4

Note(s): Signif. Codes: ^***: 0.01, ^**: 0.05, ^*: 0.1

Standard-Errors clustered by labor market (id)

All specifications at CNAE subclass level with TWFE and IVs

Source(s): Table by author

Table 3

Differences-in-differences estimates – market

Dependent variable	log(Wage)
	Simple	No covariates	Covariates	Small regions
Model	(1)	(2)	(3)	(4)
Treat x Post	0.005
	(0.004)
Treat × Year = 0		0.031***	0.033***	0.048***
		(0.006)	(0.005)	(0.008)
Treat × Year = 1		0.014***	0.026***	−0.004
		(0.004)	(0.006)	(0.011)
Treat × Year = 2		−0.006	0.008	−0.040***
		(0.005)	(0.008)	(0.015)
Treat × Year = 3		−0.013***	0.004	−0.039**
		(0.005)	(0.010)	(0.018)
Treat × Year = 4		−0.014**	0.008	−0.071***
		(0.006)	(0.011)	(0.023)
Linear trend			−0.004**	0.005*
			(0.002)	(0.003)
Municipality FE	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
Controls (demeaned and interacted)	–	–	Yes	Yes
Observations	56,668	56,668	54,883	32,766

Note(s): Clustered (Municipality) standard errors in parentheses

Signif. Codes: ***: 0.01, **: 0.05, *: 0.1

Source(s): Table by author

Table 4

Differences-in-differences estimates – merging firms

Dependent variable	log(Wage)
	Simple	No covariates	Trend	Covariates	Small regions
Model	(1)	(2)	(3)	(4)	(5)
Treat x Post	−0.006
	(0.004)
Treat × Year = 0		0.021***	0.016***	0.021***	0.030***
		(0.004)	(0.005)	(0.004)	(0.006)
Treat × Year = 1		0.026***	0.019**	0.026***	0.038***
		(0.005)	(0.008)	(0.005)	(0.009)
Treat × Year = 2		−0.029***	−0.039***	−0.029***	−0.050***
		(0.006)	(0.010)	(0.006)	(0.010)
Treat × Year = 3		−0.031***	−0.043***	−0.031***	−0.064***
		(0.006)	(0.012)	(0.006)	(0.011)
Treat × Year = 4		−0.022***	−0.037**	−0.023***	−0.058***
		(0.007)	(0.015)	(0.006)	(0.013)
Linear trend			0.002
			(0.002)
Municipality FE	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes
Controls (demeaned and interacted)	–	–	–	Yes	–
Observations	17,595	17,595	17,595	17,514	13,204

Note(s): Clustered (Municipality) standard errors in parentheses

Signif. Codes: ^***: 0.01, ^**: 0.05, ^*: 0.1

Source(s): Table by author

Table 5

Differences-in-differences estimates – merging firms and incumbent workers

Dependent variable	log(Wage)
	No covariates	Trend	Covariates	Small regions
Model	(1)	(2)	(3)	(4)
Treat × Year = 0	0.011*	0.023***	0.026***	0.022**
	(0.006)	(0.007)	(0.007)	(0.011)
Treat × Year = 1	0.009	0.027**	0.032***	0.024*
	(0.007)	(0.011)	(0.010)	(0.013)
Treat × Year = 2	−0.017**	0.007	0.010	−0.027*
	(0.008)	(0.014)	(0.014)	(0.014)
Treat × Year = 3	−0.036***	−0.006	−0.003	−0.056***
	(0.009)	(0.017)	(0.017)	(0.016)
Treat × Year = 4	−0.029***	0.007	0.010	−0.050***
	(0.008)	(0.020)	(0.020)	(0.017)
Linear trend		−0.006*	−0.007**
		(0.003)	(0.003)
Municipality FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
Controls (demeaned and interacted)	–	–	Yes	–
Observations	16,643	16,643	16,149	11,501

Note(s): Clustered (Municipality) standard errors in parentheses

Signif. Codes: ***: 0.01, **: 0.05, *: 0.1

Source(s): Table by author

Notes

1. We formally requested access to the data, but COVID-19 forced the closure of IBGE’s restricted data room.

2. A discussion about the construction of event-study models/plots and its identifying assumptions can be found in Freyaldenhoven, Hansen, Pérez, & Shapiro (2021).

Supplementary material

The supplementary material for this article can be found online.