The effects of a place-based tax cut and minimum wage increase on labor market outcomes

The decline, stagnation, and sudden rise in the value of the minimum wage in Mexico can be described in light of the economic events of the last few decades. Between 1980 and 1995, Mexico faced two adverse economic events: the 1982 debt crisis and the 1994–1995 financial crisis (better known as the Tequila Crisis). In an attempt to preserve macroeconomic stability, the government did not peg the minimum wage to the inflation rate; it lost 72% of its value and there was a substantial increase in inequality (Bosch and Manacorda, 2010). Between 1996 and 2012, the real minimum wage was kept relatively constant to avoid inflationary pressures and reflect labor productivity.

From 1987 to 2012, the government divided the country into three economic regions, each with its own minimum wage, and recent efforts to increase it have relied on reducing the number of regions: from three to two in 2012 and one in 2015. The latter reduction meant an increase of approximately 3% in some regions in that year.

Using this exogenous variation in the minimum wage by region, Campos-Vazquez et al. (2017) and Bouchot (2018) found no effect on employment outcomes. However, the increase in the minimum wage in these studies is small (less than 5%).

The increase in the minimum wage in Mexico is one of the largest recorded since 1960, according to the OECD databases, which include OECD countries and others such as Costa Rica, Brazil, Russia, and Colombia. Between 1960 and 2018, only three countries raised their minimum wage more than 50% in 1 year (in constant monetary units): New Zealand in 1973 (54%), Latvia in 1993 (103%), and Russia in 2007 (92%).

Muravyev and Oshchepkov (2015) found that the increase in the minimum wage in Russia decreased employment, especially among youth. In Mexico, however, the doubling of the minimum wage targets a specific region, which allows for a better identification.

The increase is also meaningful in terms of the ratio of the monthly minimum wage to the average monthly wage. In 2018, this ratio was 26% for formal workers on the northern border of Mexico. Among OECD countries, only the United States had a lower ratio (24% in 2017). After the minimum wage increase in Mexico, the ratio reached 45% in the border region, similar to that of Australia (46%), Chile (49%), and the United Kingdom (44%), but lower than those of Colombia (58%), France (50%), and New Zealand (52%).

All data for minimum wages in specific countries is obtained from ECLAC (2019), Eurostat (2019), and OECD (2019).

The importance of the minimum wage increase on the northern border is also in the proportion of workers affected. In 2018, about 0.1% of workers enrolled in social security (Instituto Mexicano del Seguro Social [IMSS]) earned the minimum wage. However, about 40% of workers earned between one and two times the minimum wage, the largest earnings bracket in terms of the number of minimum wages.

See Figures A1 and A2 in Supplementary Materials.

In December 2019, the government decided to increase the minimum wage 5% in the NBFZ and 20% in the rest of the country for 2020. Given this change, we focus only on the short-run effects of the policy implemented in 2019.

The VAT reduction at the northern border is not a new incentive. In the late 1970s, the Mexican government introduced the VAT at an initial rate of 10%, and in 1983 it introduced a preferential rate of 6% for consumers within 20 km of the United States, Belize, and Guatemala. In 1995 the VAT increased to 15% and 10% at the border, and in 2010 the rates were increased to 16% and 11%. In 2014, after almost 20 years of differential rates, the government unified the VAT to 16% throughout the country.

Although there have been some attempts to explore the effects of this preferential VAT rate (e.g., Davis 2011; Fuentes Flores et al. 2016), there is, to the best of our knowledge, no substantial evidence linking the preferential rate at the northern border to economic outcomes.

In January 2019, a presidential decree began to apply tax incentives to businesses and taxpayers with entrepreneurial activity residing in municipalities at the northern border. These incentives, which will remain until 2020, are a one-third discount in the CIT and a reduction in the VAT from 16% to 8%. With these measures, the government intends to restore investment, employment, and competitiveness, and to reduce crime and the high level of violence. Both incentives exclude specific sectors, taxpayers, and income sources: the financial and real estate sectors, maquiladoras, taxpayers under audit, income obtained from intangible goods or assets, e-commerce, digital content, salaries, dividends, professional services, and others (Presidencia de la República, 2018).

For businesses to take advantage of the CIT discount, they had to submit a request to the Mexican tax authority (Servicio de Administración Tributaria, SAT) and fulfill specific requirements. Businesses resident in the region for at least 18 months before the request, that generated at least 90% of their income within the region, that were not receiving any other fiscal benefit, and that had no relation with firms issuing false invoices were enrolled in a Registry of Tax Incentives for the Northern Border Region. Businesses initiating operations in the region could also request enrollment in the Registry if they could prove they would make new investments in fixed assets and facilities, and would generate at least 90% of their total income in the region.

In practice, the policy operated mainly through a reduction in the VAT and an increase in the minimum wage. By the end of 2019, among firms that applied, only 3% obtained the CIT reduced rate. In contrast, the reduction in VAT could be seen immediately through lower energy prices from the state-owned oil company Petróleos Mexicanos (PEMEX), which is the primary retailer and distributor of gasoline to private retailers. Data collected for the Consumer Price Index show that in northern cities, energy prices, including those set by the government, fell by 7% (Campos-Vazquez and Esquivel, 2020). The reduction in energy prices has had an effect on firms’ costs in the area.

This study uses administrative and household survey data to provide a fuller assessment of how a minimum wage increase coupled with tax incentives affected labor market outcomes. We employ administrative records made public on a monthly basis by the Mexican Institute of Social Security (IMSS, 2015–2019) and survey data from the National Survey on Employment and Occupations (ENOE) published quarterly by the National Institute of Geography and Statistics (INEGI, 2015–2019).

The IMSS keeps records of the number of workers registered for social security and the earnings reported by both employers and the self-employed, as required by law. This database contains mostly employees from the formal sector, whose registration is mandatory.

According to the Mexican social security law, there are two types of affiliations: the compulsory regime and the voluntary regime. Employers are required to register employees in the compulsory regime. Workers without formal employment, such as domestic workers and the self-employed, are eligible to register in the voluntary regime.

The labor force survey (ENOE), similar to the US Current Population Survey, is used in the analysis for robustness checks and to include effects not only on the formal sector, but also on the population as a whole. The sample consists of people aged 15 or older in randomly selected households at the national level, and it is representative at the national and state levels, and for some urban areas. The data are published on a quarterly basis and each household in the sample is interviewed for five consecutive quarters. The ENOE provides information about the main socioeconomic and demographic characteristics of the Mexican labor force, with detailed information concerning earnings, occupation, economic activity, and working conditions.

Both datasets have advantages and disadvantages. The main advantage of the IMSS sample is its extensive coverage, which allows the measurement of formal employment and earnings at the municipality level on a nationwide, monthly basis. However, employers might have an incentive to report fewer employees or lower salaries to evade payment of their social security contributions. If this incentive varies with the economic stimulus package, the labor force survey may find different estimates. Therefore, incorporating the ENOE sample into the analysis provides a robustness check and further support for the results observed with the IMSS sample.

Table 1 shows the main descriptive statistics using IMSS data for municipalities in the border and non-border regions. There are 18 municipalities (those with at least 50,000 inhabitants) in the treatment group and 395 in the control group. The border region represents close to 11% of total employment. The proportion of men in the formal sector is higher than that of women, but the border region employs relatively more women than the non-border region. Most of the formal sector employment is generated by large firms, especially in the border region. Less than 3% of workers are 60 years of age or older. The proportion of workers earning less than MXN 180 daily is higher in the non-border cities. Approximately 31% earned less than MXN 180 daily in the border municipalities in September: these are the workers directly affected by the new minimum wage. The average wage for workers in the border region is USD 129.4, so the new minimum wage means an average increase of 36%. The average monthly labor income for all workers is MXN 9,427 (approximately USD 1,050 in PPP).

In Supplementary Material, we include IMSS data showing that in 2018 most formal sector workers in the country earned between one and two minimum wages, the same income as approximately 40% of all workers; this share increased to 43.8% in 2019. Of this share, only 0.1% earned only one minimum wage in both years (Figure A1). According to the ENOE, in 2018 the proportion of workers earning at most one minimum wage was about 16.1%—3.3% in formal and 25.9% in informal employment.

This study focuses on the short-run effect on labor market outcomes of a stimulus package of location-based policies in the Mexican northern border region. Since the policy affected only this region, which includes a specific set of municipalities, it is possible to identify the impact of the intervention. However, it is necessary to compare it with a group of municipalities that have had the same secular trends in the outcome variables, but did not receive the stimulus. To build such a group, we employ the SCM, which relies on a set of predictors, in our case pre-intervention outcomes that are used to assign weights to unaffected municipalities, such that the trends of the weighted average of pre-intervention outcomes best represent the trends of the counterfactual of the treated unit.

Following standard notation (e.g., Abadie et al., 2010; Cavallo et al., 2013), we consider J+ 1 municipalities. Without a loss of generality, the first unit is a municipality in the northern border region, where the stimulus package was implemented in January 2019. The remaining J municipalities constitute the donor pool, which is the set of municipalities that did not receive any benefit from the stimulus package. Then, the effect of the intervention on the outcome variables for the treated municipality, for example, formal employment, E_it, can be written asfollows:

where E1tN$E_{1t}^{N}$is the outcome that would be observed in the affected municipality in the absence of the intervention and E1tl$E_{1t}^{l}$the outcome in the same region with the minimum wage increase and tax incentives. Since E1tl$E_{1t}^{l}$is observed, to estimate the effect on employment we need to estimate E1tN.$E_{1t}^{N}.$The SCM constructs a vector of weights W over J donor municipalities, such that the weighted combination closely matches the treated municipality in pre-intervention formal employment.

Following Abadie and Gardeazabal (2003), Abadie et al. (2010), and Abadie et al. (2015), we proceed to define the optimization problem to compute the synthetic control. Formal employment (or earnings) is observed for T periods for the Mexican northern border region (or a municipality within a region) E_1t, and for the remaining unaffected municipalities E_jt for t = 1,…, T and j = 2,…, J + 1. Intervention happens at time t = t′ (January 2019). We define a vector of pretreatment characteristics for the treated region (4 years of monthly observations, or up to 48 observations for each characteristic) formed only by pre-intervention outcomes or predictors X1=(E1,t′−48,…,E1,t′−1).${{X}_{1}}=\left( {{E}_{1,{t}'-48}},\ldots ,{{E}_{1,{t}'-1}} \right).$Similarly, we define a matrix X0=(Ej,t′−48,…,Ej,t′−1)${{X}_{0}}=\left( {{E}_{j,{t}'-48}},\ldots ,{{E}_{j,{t}'-1}} \right)$containing the same pre-intervention outcomes or predictors for the J donor municipalities. Defining a (J × 1) vector of positive weights (J × 1) whose sum is 1, with W = (w₂,…, w_{J+ 1}) as the weighted average of available donors, the following optimization problems determine the optimalweights that describe the synthetic control:

where n is a symmetric and positive semidefinite matrix whose elements reflect the importance of each predictor according to its predictive power for the outcome variable. Given the optimal weights W=(w2⋆,…,wj+1⋆)$W=\left( w_{2}^{\star },\ldots ,w_{j+1}^{\star } \right)$for each donor, the synthetic control at any time t is the weighted combination of donor employment ∑j=2J+1wj⋆Ejt;$\sum\limits_{j=2}^{J+1}{w_{j}^{\star }{{E}_{jt}}};$the estimate of the effect on employment is therefore:

The optimization problem covers the period from January 2015 to December 2018 with monthly observations of formal employment (or earnings), and we estimate the effect on formal employment for the period from January to December 2019. Although it is clear how to compute matrices n and W, there is little guidance and a lack of consensus about the predictors used to construct a valid synthetic control other than a good pretreatment fit (Abadie et al., 2010), which runs the risk of specification searching to find significant estimates (Ferman et al., 2017). We therefore allow the algorithm to choose the best fitting model. Like Abadie et al. (2010), we use the minimum value of the root mean square prediction error (RMSPE) as the key criterion in the pretreatment period. To find the best model for each outcome variable, we consider a set of up to 50 different specifications to determine which model meets this criterion. These specifications include lags of the outcome variable and covariates that do not vary over time (like the poverty head count ratio and other variables described in Table 1, as well as average labor income in different periods). We report as the main results the output produced by the model with the minimum RMSPE.

An additional complication in the analysis is that we have 18 municipalities as treated units. We thus estimate 18 different synthetic controls, one for each municipality, and obtain a treatment effect for each municipality and outcome variable. To obtain the aggregate effect for the northern border region, we use a weighted average using population in treated municipalities in 2015 as weights. The results are similar if we use one aggregated treated unit instead of the 18 individual municipalities and build its synthetic control accordingly (Figure A6 in Supplementary Materials).

To make inferences from this estimate, we provide two results. First, we follow Abadie et al. (2010, 2015) and Cavallo et al. (2013) and show p-values using placebo tests, calculated by repeating the synthetic control procedure for each donor unit excluding the treatment units. The p-value is obtained as the proportion of cases in the placebo tests whose effects are as large as the estimated treatment effect.

We use the Stata synth_runner package to implement this procedure (see Galiani and Quistorff 2017). The p-value is calculated for those placebo tests with a certain value of RMSPE to ensure that they are not noisy and that good controls are found for each placebo. We modify the package to calculate p-values using a stricter criterion for the RMSPE to ensure that the p-values are not sensitive to bad controls.

To implement the SCM, we normalize the value of all outcome variables (employment and earnings) in September 2018 for each municipality. This allows us to compare employment and earnings trends across different municipalities and increases the number of potential donors who can be included in the synthetic control group. For example, the method can consider donors with trends similar to those in the treatment group regardless of their level of employment.

A related approach, similar to that of Card (1992), is to estimate the elasticity of labor demand using the local stimulus package in the northern border region as the instrumentalvariable. The regression to estimate is:

where the dependent variable is the index of employment level (E_mt) and the main independent variable is the index of the earnings level (W_mt). Parameters δ_m and δ_t refer to municipality and year–month fixed effects. The first controls are for unobserved characteristics that are constant over time for each municipality, and the second controls are for effects that affect all municipalities at the same time (like macroeconomic effects that vary seasonally throughout the country). The matrix X_m contains control variables that do not vary over time, like the poverty rate in 2015 and other variables shown in Table 1. The variable Fraction_m,2015 refers to the fraction of workers in 2015 earning less than MXN 180 (close to the new minimum wage enacted in 2019 for the northern border region) in municipality m. Standard errors are clustered at the municipality level. The instrumental variable for W_mt is a dummy variable with the value of 1 if the municipality is in the northern border region and observed in 2019, the period of the local stimulus package. We show regressions using all municipalities and using municipalities in the synthetic control group (in which case we include only time fixed effects and treatment group with no further controls).

To ensure the robustness of the results using the SCM, we implement other strategies and econometric methods. First, the SCM relies on finding a valid counterfactual. It is possible that the northern border in Mexico depends much more on the United States than on the Mexican economy. We calculate the SCM using only metropolitan areas in the United States from the Current Employment Statistics of the US Bureau of Labor Statistics. Second, we obtain the treatment effect using the procedure described in Doudchenko and Imbens (2017). Finally, to corroborate the results using the full population we rely on the labor force survey (ENOE) for the same period as before: 2015–2019. The dataset has various disadvantages: the sample does not cover all municipalities, and labor income suffers from measurement error and a low response rate (Campos-Vazquez and Lustig, 2017). However, the main advantage is that it covers the full working population, including the formal and informal sectors, and includes information at the individual level. This allows us to include individuals’ age, gender, schooling, and marital status as control variables.

In this section, we describe the results from using the labor force survey (ENOE) for the period 2015–2019. The advantage of this dataset is that we can analyze patterns for total employment including the informal sector. We focus on the following outcome variables (for individuals 20–65 years old): the employment rate (number of employed individuals divided by the population), the female employment rate, the share of workers in the formal sector (Formality (Emp)), the share of formal sector workers in the population (Formality (All)), labor income as observed and using imputation to avoid the low response rate, and school attendance (for individuals 15–24 years old).

We use a hot-deck procedure to impute labor income when there is no response. The imputation is within year, quarter, gender, age groups, schooling groups (less than primary, primary, junior high school, high school, and college), formal sector status, state, and border region. We calculate the average labor income across five imputations.

In this case, it is not possible to use the synthetic control methodology, as the sample size within municipalities is too small to calculate trends at that level. We rely on simple OLS regressions using the same control variables as earlier, but with the unit of observation at the individual level. This allows us to also include controls like individual age, gender, schooling, and marital status. The key regressor is a dummy variable for the border region in 2019. The results include the estimation of individual control variables only, and also with controls at the municipality interacted with a trend, with state-level trends, and finally municipality fixed effects interacted with a trend. The results are plotted in Figure 7 (results for different subsamples and restricting for the period 2017–2019 are shown in Figures A10–A16 in Supplementary Materials).

For example, we estimate models for different age groups, for workers with less than high school, and for workers in the cities selected by the synthetic control method. All results are similar.

Figure 7

The results are broadly consistent with the previous findings using IMSS administrative data. There is a small negative effect on total employment (using state trends or municipality trends), although the estimate is not statistically significant and its magnitude is relatively small. There is a small nonsignificant negative effect on female employment (consistent with the finding that the elasticity of labor demand is negative for women, especially in small firms). In contrast with the results using administrative data, when all control variables are included there is a positive although non-statistically significant effect on formality (rows with the variables Formality (Emp) and Formality (All)), suggesting that employment decreases mainly through a decrease in informality.

To corroborate this claim, we estimate transition probabilities using a multinomial logit, taking advantage of the panel structure in the labor force survey. For the period 2015–2019, we take the observations for the first and fifth quarters for the same individual, and estimate a multinomial logit conditioned on being out of the workforce, being in the formal sector, or being in the informal sector in the first quarter. The model thus estimates the transition probability 1 year later of being out of the workforce, having formal employment, or having informal employment. The key regressor is a dummy variable for the northern border cities interacted with a dummy variable for the year 2019. The results clearly indicate that the transition from formal to informal employment is reduced and the transition from informal to formal employment increases. The other effects are close to zero. See Figure A17 in Supplementary Materials.

The findings using administrative data suggest a null to small negative effect on total employment due to the place-based policy. But using the variation of wages within demographic groups, the findings suggest that the elasticity of labor demand is negative and significant for small firms, and especially for women.

Using household survey data, the findings suggest that the estimates for total employment and those for the proportion of formal sector workers in the total population are close to zero. The estimates are slightly more negative for women, but they are still nonsignificant. In sum, labor income increases substantially for workers in the northern border municipalities, with no large effects (positive or negative) on employment outcomes.

Both datasets suggest that if there is a negative effect of doubling the minimum wage, it is likely to be relatively small at the aggregate level. To understand these results, we search for all newspaper articles published in or mentioning the border municipalities that use the term “minimum wage.” There were 39 articles from January to September 2019, and only 6 were about the potential negative effects of the minimum wage (mainly in January and February). However, there were no articles documenting firm closures or employment contractions due to the new minimum wage. On the contrary, recent articles document how Central American migrant workers are being hired in the maquila and restaurant industries at the new minimum wage.

Voz de América, June 28, 2019, “México ofrecerá 40.000 empleos a migrantes en su frontera con EE.UU”; Los Angeles Times, September 4, 2019, “Ante una larga espera en Tijuana algunos migrantes encuentran trabajo”; El Mañana, September 27, 2019, “Contratan restauranteros a la comunidad migrante”; El Mañana, September 29, 2019, “Dan trabajo maquilas a unos 120 migrantes.”

One explanation for the virtually null effect of the policy on employment is that firms might be offsetting labor costs through higher prices. With the simultaneous decrease in VAT and increase in the minimum wage, real wages are likely to rise along with aggregate demand. In fact, analyzing previous VAT reforms in Mexico, Mariscal and Werner (2018) found that the pass-through from VAT to inflation is approximately 20%. A potential decrease in inflation in the northern border municipalities is thus about 1.4% points (VAT cut from 16% to 8%, 1.08/1.16 × 20%). In 2018, average annual inflation in both the border region and the rest of the country was 4.9%, while in September 2019 it was 2.4% and 4%, respectively.

Results are approximately the same using a regression framework. Campos-Vazquez and Esquivel (2020) obtained a similar conclusion.

The government also introduced other new programs in January 2019. The most relevant is the program “Young People Building the Future” (“Jóvenes construyendo el futuro”), which provides a monthly stipend of MXN 3,600 (approximately USD 360 in PPP) to individuals 18–29 years old who are not currently employed and are willing to work as apprentices. In the IMSS administrative data they are not counted as employed, since the government, not the employer, pays their social security contributions. Their presence in the labor force survey depends on how they interpret the question “During the past week, did you work at least one hour?” But even if they answer no, there is a filter for training courses that considers the person employed. Hence, IMSS data do not take into account the individuals in the program, but the labor force survey does.

The main results using administrative data show that restricting for age groups does not change this null to small negative effect. When we estimate all effects, omitting individuals younger than 30, the results are approximately the same. Moreover, for the youth program to bias the estimates, enrollment would have to be larger in the northern border municipalities. However, using administrative data we find that enrollment in these municipalities is close to 70 apprentices per 100,000 inhabitants, while in the rest of the country it is close to 600. The youth program thus does not seem to be affecting the main results of the study.

It is possible that the lack of an employment effect is due to a compensation of costs from the VAT reduction and the increase in labor costs. This is difficult to ascertain, as not all of the information needed to compute different scenarios is available. The Mexican Economic Census is available for information on costs for all firms in 2018. In the NBFZ, labor costs account for 26.9% of total costs, energy costs account for 5.3%, and other costs for 67.8%. We also need an estimate of how prices change. Previous results indicate that labor income increased by 9%. Campos-Vazquez and Esquivel (2020) show that energy prices fell by 7%, and core inflation (which excludes energy prices) fell by 0.9%. With quantities fixed, labor costs increase by 2.4% and energy costs decrease by 0.4%. The rest is difficult to estimate, as we do not know how much of inputs is bought in local markets (with a VAT of 8%) as opposed to the rest of the country (with a VAT of 16%). If we assume, for example, that only 10% of nonlabor, non-energy costs is due to inputs bought locally (and 57.8% from those bought in the rest of Mexico or abroad), we obtain a reduction in total costs of 0.09%. Under this scenario, total costs increase by 1.9%. From this calculation, it is clear that the policy has heterogeneous effects, depending on an establishment’s share of labor, energy, and nonlabor and non-energy costs on total costs, as well as on how much of inputs is bought locally (and subject to lower prices). In any case, the reduction in VAT partially or completely compensates for the increase in labor costs. Matched employer–employee data might provide a clearer picture on how costs were modified under the policy and its effects on employment outcomes at the individual and establishment level.