**Hypothesis 1.** *The effect of Chinese import penetration has a greater magnitude than that of ROW import penetration*.

Moreira (2006) points out that China's accession to the WTO exposed its trade partners to an unparalleled trade shock due to the uniqueness of the labor abundance of the Chinese economy relative to almost all countries in the world (Brazil included). Hence, according to the Heckscher-Ohlin model, unskilled-labor intensive imports from China apply a stronger competitive pressure than other imports. This is supported by Ashournia et al. (2014), who found that low-skill intensive firms in Denmark were heavily impacted by Chinese imports. Conversely, the competitiveness of ROW imports is milder in unskilled-labor intensive industries relative to that in other industries. This leads to the second testable prediction.

**Hypothesis 2.** *Increased Chinese import penetration has a larger effect on unskilled-labor intensive industries relative to its impact on other industries. Conversely, increased ROW import penetration affects other industries more than unskilled-labor intensive industries*.

Another factor that can modulate the effects of import competition on labor market outcomes is the location of manufacturing activity. Indeed, this exhibits a significant spatial heterogeneity in Brazil, as 7 out of 26 states account for approximately 80 percent of all manufacturing activity (Paz 2019b). These seven states are those that make up the Southeastern and Southern regions of Brazil. According to Kapri and Paz (2019), such a spatial concentration makes manufacturing workers experience a different exposure to trade-induced shocks depending on their location. Thus, the effects of import penetration are expected to be different on workers in states with a large manufacturing sector. This leads to the third and last testable hypothesis:

**Hypothesis 3.** *Increased industry-level import penetration has a distinct impact on workers in manufacturing states relative to those in non-manufacturing states*.

#### **3. Empirical Methodology**

The empirical methodology used to assess the testable hypotheses exploits the industrylevel variation in competition induced by imports from China and from the ROW on the worker-level informality likelihood and on average formal and informal wages. Paz (2014) points out that decisions about the type of labor contract used (formal or informal) and the wage paid are made simultaneously. Hence, overlooking this simultaneity leads to biased estimates of the effects of import competition on wages. This was found to be the case in Brazil during the 1990s by Paz (2014) and in Peru during the 2000s by Pierola and Sanchez-Navarro (2019). Paz (2014) proposes addressing this simultaneity by means of the two-step switching regression framework from Maddala (1983). The first step in this

framework is the regime selection equation that models the choice of the employment type using a Probit discrete choice specification, as depicted by Equation (1).

$$\begin{array}{c} \text{infformal}\_{ij \pm t} = a + \beta\_1 I P\_{j, \ t-1}^{\text{Clina}} + \beta\_2 I P\_{j, \ t-1}^{\text{ROW}} + \Psi \chi\_{ij \text{st}} + \text{other} \\ \quad + \theta\_s + \delta\_t + u\_{ij \text{st}} \end{array} \tag{1}$$

 $\text{[information]}\_{\text{ijst}}^{\*} = 1$   $\text{if } \text{[informal}\_{\text{ijst}} > 0, \text{ } \text{[informal}\_{\text{ijst}}^{\*} = 0 \text{ } otherwise.]$ 

where *inf ormal*∗ *ijst* is an indicator dependent variable that is "1" if worker *i* in industry *j* in state *s* and year *t* holds an informal job, and is "0" if worker holds a formal job; and *inf ormalijst* is the respective latent variable. *IPChina <sup>j</sup>*, *<sup>t</sup>*−<sup>1</sup> is the Chinese import penetration. *IPROW <sup>j</sup>*, *<sup>t</sup>*−<sup>1</sup> is the ROW import penetration. <sup>X</sup>*ijst* is a vector of worker *<sup>i</sup>*'s observable characteristics, namely age, age squared, number of years of education, and indicators for gender, race, marital status, high school degree, and college degree. *γ<sup>j</sup>* are industry fixed effects that control for the industry-specific and time-invariant characteristics. *θ<sup>s</sup>* are state fixed effects which capture state specific and time-invariant characteristics, such as being landlocked. *δ<sup>t</sup>* are year effects that account for time-varying factors that affect industries equally, such as business cycles. *ujt* is the error term.

For identification purposes, the switching-regression framework requires the selection equation to contain at least one relevant variable that does not affect the wage earned by the worker. Following Paz (2014), the variable fulfilling this role is *otherformalijst*. This is a dummy variable that is "1" if another person in the household has a formal job, and "0" otherwise. The idea is that the tradeoff between a formal and an informal job experienced by a worker is affected by another household member having a formal job, since informal labor contracts are used by both firms and workers to evade taxation. In this vein, having another household with a formal job may increase the likelihood of detection of income tax evasion, and this reduces the incentives for choosing an informal job. This suggests a negative coefficient for this indicator variable in the selection equation.

The second step of the switching regression framework is to model the average wage through a Mincer-type wage (*yijst*) equation for each type of job using the inverse Mills ratio to control for the worker self-selection into that regime. The inverse Mills ratio (Λ(*z*)) is defined as Λ(*z*) ≡ φ(*z*)/Φ(*z*), where φ(*z*) is the standard normal distribution function and Φ(*z*) is the standard normal cumulative distribution function. It is calculated using the predicted values of the informality likelihood (*inf ormal ijst*). These wage equations are depicted by Equation (2) for the formal worker and by Equation (3) for the informal worker.

$$y\_{ijst}^{for} = a^{for} + \beta\_1^{for} IP\_{j\_s \ t - 1}^{China} + \beta\_2^{for} IP\_{j\_s \ t - 1}^{ROW} + \Psi^{for} \mathbf{X}\_{ijst} + \gamma\_j^{for} + \theta\_s^{for} + \delta\_t^{for} \tag{2}$$

$$+ \sigma^{for} \Lambda \left( - \widehat{
ormalfont
\,
} \right) + \upsilon\_{ijst}^{for} \tag{3}$$

$$y\_{ijst}^{inf} = a^{inf} + \beta\_1^{inf} I P\_{j\_s \ t - 1}^{China} + \beta\_2^{inf} I P\_{j\_s \ t - 1}^{ROW} + \Psi^{inf} \chi\_{ijst} + \gamma\_j^{inf} + \theta\_s^{inf} + \delta\_t^{inf} \tag{3}$$

$$+ \sigma^{inf} \Lambda \left( \widehat{normal}\_{ij \text{st}} \right) + v\_{ij \text{st}}^{inf} \tag{3}$$

The selection into an informal or a formal labor contract takes place as a non-zero correlation among the error terms *uijst*, *<sup>v</sup>f or ijst*, and *v inf ijst*. Should this be the case, the estimated coefficients of the inverse Mills ratios will be statistically significant. Accordingly, the omission of these terms will then lead to biased estimates of the effects of the import penetrations on wages. Note that the standard errors for Equations (2) and (3) are estimated by means of a 500-repetition bootstrap using household survey weights because the inverse Mills ratio is a generated regressor.

The calculation of the marginal effects of the import penetrations on wages ought to consider the effects of the import penetrations on the regime selection. This means that the effects of the import penetrations on wages will, in a non-trivial way, depend on the workers characteristics via the selection equation. The marginal effect of Chinese import penetration on the formal and the informal wages are given by Equations (4) and (5), respectively. The marginal effects for ROW import penetration are calculated in a similar fashion.

$$\frac{\partial \text{wage}^{\text{for}}\_{\text{ijst}}}{\partial \text{IP}^{\text{Cl}}\_{\text{j}, \text{t}}\_{\text{t}} - 1} = \widehat{\beta\_1^{\text{for}}} + \widehat{\sigma^{far}} \Lambda \left( - \inf \widehat{\text{normal}}\_{\text{ijst}} \right) \widehat{\beta\_1} \left[ \inf \widehat{\text{normal}}\_{\text{ijst}} - \Lambda \left( - \inf \widehat{\text{normal}}\_{\text{ijst}} \right) \right] \tag{4}$$

$$\frac{\partial wage\_{ijst}^{inf}}{\partial IP\_{j,t-1}^{China}} = \widehat{\beta\_1^{inf}} - \widehat{\sigma^{inf}} \Lambda \left( \inf \widehat{formal\_{ijst}} \right) \widehat{\beta\_1} \left[ \inf \widehat{normal\_{ijst}} + \Lambda \left( \inf \widehat{normal\_{ijst}} \right) \right] \tag{5}$$

To assess the predictions of Hypothesis 2, Equations (1)–(3) are augmented to include new regressors that are interactions between the import penetration measures and *LIj*, which is an indicator variable that is "1" for unskilled-labor intensive industries, and "0" otherwise. As in Paz and Ssozi (2021), the unskilled-labor intensive industries are the seven industries with the lowest average of year of schooling in 2000, namely food and beverages, textiles, apparel, footwear and leather products, wood products, non-metallic minerals and products, and furniture and other products. Similarly, Hypothesis 3 is evaluated through the addition to Equations (1)–(3) of interaction terms between the import penetrations and an indicator variable (*Manufacturing states*) that is "1" if state *s* is in the South or Southeast of Brazil, and "0" otherwise.

There are some concerns about the empirical methodology that deserve careful discussion. First, Brazilian producers may react slowly to changes in market conditions. This can be addressed by employing lagged import penetrations. The second concern is that both the import penetration measures and wages are simultaneously determined because the latter is part of the value added that is used in the calculation of the import penetration. This concern can also be alleviated by using the first lag of the import penetration measures.

The third concern is the omitted variable bias. More precisely, this could be the case of omitted time variant industry-specific shocks that affect both import penetrations and outcomes, and this biases the estimates. Examples of such omitted variables are demand or supply shocks in the Brazilian economy. For instance, suppose that a larger than expected import penetration increase leads the Brazilian government to impose higher import tariffs, safeguards, or countervailing duties. This behavior can be seen in the number of antidumping procedures initiated in Brazil that reached almost 100 cases in the 2000s, of which a quarter were against Chinese producers (cf. WTO Antidumping Gateway 2016). These product-level non-tariff protection measures cannot be accounted for in the empirical specification, and thus can bias the estimates. The use of an instrumental variable (IV) approach can address this concern, as explained next.

The excluded instruments used in the IV strategy are based upon Iacovone et al.'s (2013) idea of using supply-driven shocks as an instrument for import penetration.2 Thus, the excluded instrument for Chinese import penetration is the Chinese share of imports in third countries. The third countries considered have very small trade ties with Brazil and are located in Latin America, namely Mexico, Colombia, Costa Rica, Ecuador, El Salvador, Guatemala, Guyana, Jamaica, Nicaragua, Panama, and Peru. The correlation between the Chinese import penetration and Chinese share of imports in third countries is 0.574. The same endogeneity concern applies to the ROW import penetration; therefore, an additional excluded instrument is required. In a similar fashion, this is the high-income countries' share in the imports of the above mentioned third countries. These high-income countries are Australia, Austria, Belgium, Bulgaria, Canada, Croatia, Czechia, Denmark, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Italy, Japan, Luxembourg, Netherlands, Norway, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, Sweden, Switzerland, USA, and United Kingdom. The simple correlation between the ROW import penetration and the high-income share of imports is 0.316. In the specifications that include interaction terms between the import penetrations and indicator variables, the additional

excluded instruments are built by interacting the excluded instruments mentioned above with the respective indicator variable.

Table 3 reports the estimated coefficients of the regressions of the endogenous explanatory variables (Chinese and ROW import penetrations) on the excluded instruments and other control variables used in the selection Equation (1). We can see that the Chinese share of imports in Latin America is statistically significant when the Chinese import penetration is the dependent variable, column (1), while the other excluded instrument is not significant. Similarly, the high-income countries' share of imports in Latin America is significant in the specification for the ROW import penetration in column (2), whereas the Chinese share of imports in Latin America is not significant in this regression. None of the workers' characteristics are statistically significant. These estimates show that the excluded instruments have predictive power over the endogenous regressors. The *F*-statistic of these regressions in Table 3 is above 200, which suggests that a weak instrument is not the problem.


**Table 3.** First-stage OLS regressions of the endogenous regressors.

Number of observations is 671,134. \*\*, and \* indicate statistical significance at the 5%, and 10% levels, respectively. Standard errors clustered at industry level are reported in parenthesis. Household survey weights used.

#### **4. Results and Discussion**

This section begins with the estimates of the selection specification based on equation (1), which is the first step of the switching regression framework. Table 4 displays the estimated coefficients obtained using Probit in columns (1)–(3) and those obtained using IVProbit in columns (4)–(6). The worker's characteristics' estimated coefficients do not vary across specifications. They indicate that older, female, married, and Asian workers are more likely to hold an informal job, while black workers are less likely. The likelihood of having an informal job is smaller for those with a high school degree and with a larger number of years of schooling. Interestingly, the college degree indicator is not statistically significant at the 5 percent level in any specification. The other household member has a formal job indicator estimated coefficient which is negative and statistically significant in all specifications of Table 4, with a similar magnitude across specifications.



Notes: Number of observations is 671,134. \*\*\*, \*\*, and \* indicate statistical significance at the 1%, 5%, and 10% levels, respectively. Standard errors clustered at the industry level. Sample weights from PNAD/Census used. State, industry, and year fixed effects included in the estimated model. The excluded instruments are the Latin American countries' Chinese share of imports, the Latin American countries' high-income countries share of imports, and their interactions.

The Probit specifications do not display statistically significant coefficients for either the import penetrations or for their interactions. Additionally, notice that the null hypothesis of exogeneity of the endogenous regressors is rejected at the 5 percent level in all IVProbit specifications. This suggests that the focus of the analysis should be on the IV Probit specifications since they account for the endogeneity of the import penetrations and their interactions.

The IVProbit specification of column (4) displays no statistically significant coefficient of the Chinese and the ROW import penetration, therefore this result renders no support for Hypothesis 1. The specification in column (5) is designed to assess Hypothesis 2. It presents the Chinese and the ROW import penetrations with positive and statistically significant coefficients. The interaction terms of the import penetrations with the unskilled-labor intensive indicator are negative for both import penetrations, and statistically significant at the 5 percent level for the Chinese and at the 10 percent level for the ROW import penetration. These coefficients imply that an increase in any import penetration raises informality in non-unskilled-labor intensive industries. For unskilled-labor intensive industries, Chinese import penetration has no effect on informality, while ROW import penetration reduces informality in unskilled-labor intensive industries. These results do support that the Chinese and the ROW import penetrations have different impacts on informality (Hypothesis 1), and such impacts also depend on the unskilled-labor intensity of the industry. These results are not in line with Hypothesis 2. The estimates to assess Hypothesis 3 are in column (6), where both import penetrations are positive and significant as before. This means that increased import penetration leads to more informality in non-manufacturing states. The interaction terms with manufacturing state indicators are negative and significant. Accordingly, for manufacturing states Chinese import penetration reduces (albeit with a smaller magnitude) informality, whereas ROW import penetration has a positive effect. These results do support Hypothesis 3.

The results for ROW imports are somewhat distinct from those obtained by Paz (2014) using Brazilian data for the 1990s. His results indicate that greater imports—most from high income countries—increased the manufacturing informality share. On the one hand, Pierola and Sanchez-Navarro (2019) had IVProbit results of no effect regarding Chinese imports on informality for Peru in the 2000s, in line with this study findings. On the other hand, their findings that Chinese imports do increase informality among unskilled workers are at odds with the results of Table 4.

Table 5 reports the OLS and IV estimates employing equation (2) for the formal workers' average wage in columns (1)–(3) and (4)–(6), respectively. The selection terms (the inverse Mills ratios) were computed using the IVProbit estimates of the specification that contained the same explanatory variables (except for the other formal indicator). The estimated coefficients of the selection terms are statistically significant in all columns. These coefficients are positive except in column (5). This is evidence that selection effects are taking place and should not be overlooked. The estimated coefficients for workers' characteristics are stable across specifications. Formal workers' wages are increasing with age and number of years of schooling, with considerable premium for high school and college degrees. Females and blacks earn a lower wage on average. Additionally, married and Asian workers have higher average wages. Notice that the OLS and the IV estimates for some variables are substantially different. In fact, the null of exogeneity of the import penetrations (and their interaction terms) is rejected at the 5 percent level in every column. Hence, the discussion will focus on the IV specifications results.

In column (4), the estimated coefficients of the Chinese import penetration are positive and statistically significant at the 5 percent level of confidence and the coefficient for the ROW import penetration is zero. As discussed in the previous section, the estimated coefficients of the import penetrations and their interactions are different from their marginal effects on the average wage due to the selection effect. The marginal effect is calculated for the average formal and informal worker, i.e., the estimated selection equation fitted at the average value of its regressors, as shown in Equation (4). The effects on the average formal wage of a percentage point increase in the Chinese import penetration is a 1.09 percent increase and for the ROW import penetration a 0.06 percent increase, respectively. These figures suggest that the Chinese and ROW import penetrations had a distinct impact on formal workers' average wage, which supports Hypothesis 1.

The specification in column (5) shows that both the Chinese and the ROW import penetrations are positive and significant. The estimated coefficients of the interaction terms with the unskilled-labor intensive indicators were negative, though only the interaction with the ROW import penetration is significant. This means that the total effect of Chinese import penetration on the formal workers' average wage for unskilled-labor intensive industries is positive, while the effect for the ROW import penetration in negative. For non-unskilled-labor intensive industries, the marginal effect on the average formal wage of a percentage point increase in the Chinese import penetration is a 3.47 percent increase, and for the ROW import penetration a 3.04 percent increase. For unskilled-labor intensive industries, these marginal effects are a 2.01 percent increase and a 4.28 percent decrease, respectively. These results are at odds with the predictions of Hypothesis 2. These results for average informal wages are at variance with the findings of Paz (2014) and Pierola and Sanchez-Navarro (2019). The former found a negative effect of ROW imports, and the latter found no effect of Chinese imports for the average worker, though their results for unskilled workers show a negative impact.

**Table 5.** Worker-level IV estimates of the effects of industry-level import penetration on the wages of formal workers using Equation (2).


Notes: Number of observations is 533,584. \*\*\* and \*\* indicate statistical significance at the 1% and 5% levels, respectively. Standard errors are bootstrapped with 500 repetitions. Sample weights from PNAD/Census used. State, industry, and year fixed effects included in the estimated model.

The estimates in column (6) display a positive coefficient for the Chinese import penetration and a negative for ROW import penetration. Both coefficients are statistically significant. The estimated coefficients of the interaction terms between the manufacturing state indicator and import penetrations are positive and significant. This means that, for manufacturing states, the total effects of both import penetrations are positive. For a formal worker located in a non-manufacturing state, the marginal effect on the average formal wage of a percentage point increase in the Chinese import penetration is a 0.1 percent decrease and for the ROW import penetration a 2.66 percent decrease. In manufacturing states, in contrast, these marginal effects are a 2.6 percent increase and a 0.64 percent increase, respectively. These estimates corroborate Hypothesis 3.

Table 6 reports the estimates for the average informal wage based on Equation (3). The OLS estimates are displayed in columns (1)–(3), and the IV estimates in columns (4)–(6). As in the previous table, the selection term estimated coefficients indicate the presence of selection effects. They are statistically significant in every specification, except in column (5). The estimated coefficients of the workers' observable characteristics are stable across specifications. They present signs that are comparable to those in Table 5; however, the coefficients for female and years of schooling have a larger magnitude, while the coefficients for Asian, high school, and college indicators exhibit a smaller magnitude. As before, the null of exogeneity of the import penetrations is rejected at the 5 percent level in the IV specifications.

**Table 6.** Worker-level IV estimates of the effects of industry-level import penetration on the wages of informal workers using Equation (3).


Notes: Number of observations is 136,382. \*\*\*, \*\*, and \* indicate statistical significance at the 1%, 5%, and 10% levels, respectively. Standard errors are bootstrapped with 500 repetitions. Sample weights from PNAD/Census used. State, industry, and year fixed effects included in the estimated model.

We can see in column (4) that the only statistically significant import penetration coefficient is for the ROW, which is positive. The marginal effect on the average informal wage of a percentage point increase in the Chinese import penetration is a 0.52 percent increase and for the ROW import penetration is a 3.39 percent increase. These figures support Hypothesis 1. The estimates in column (5) for both the Chinese and the ROW import penetrations are positive and significant at the 10 percent level. The coefficients of the interaction terms with the unskilled-labor intensive indicator are negative and not significant. Focusing on non-unskilled-labor intensive industries, the marginal effect on the average informal wage of a percentage point increase in the Chinese import penetration is a 5.4 percent increase and for the ROW import penetration a 5.3 percent increase. For the unskilled-labor intensive industries, these figures are a 0.12 percent increase and a 2.26 percent increase, respectively. These effects are at variance with the predictions of Hypothesis 2. These results for average formal wages are comparable to those of Paz (2014), who found a positive effect of ROW imports. Yet, Pierola and Sanchez-Navarro (2019) found no effect of Chinese imports on the average informal worker, and a negative effect for the unskilled informal worker.

The estimates in column (6) reveal that both import penetrations are negative, albeit not statistically significant. Their interactions with the manufacturing state indicator are positive, but only the interaction with the ROW import penetration is statistically significant. The marginal effect on the average informal wage of a percentage point increase in the Chinese import penetration is a 4.72 percent decrease, and for the ROW import penetration a 6.61 percent decrease for non-manufacturing states. The effects for manufacturing states are a 0.14 percent decline and a 2.23 percent increase, respectively. These different impacts on the average informal wage support Hypothesis 3.
