Data

The research is carried out using an approach focusing on federal states for every quarter between 1997 and 2019, the period in which most information is available. At this point, it is important to underline two aspects: first, for some variables, such as education, the methodology is estimated on an annual basis due to the scarcity of statistical information; second, the data includes municipal-level variables to complement and corroborate the meaning of the results, as well as to expand the results to a local level of disaggregation.

The statistical information regarding the automotive sector was obtained from the five-yearly economic censuses and the administrative registry of the automotive industry, belonging to the INEGI.¹In turn, data from the Mexican Automotive Industry Association (AMIA) is used.² These sources provide information regarding production, income, wages, intermediate consumption, exports, manufacturing companies and FDI. Meanwhile, employment and salary statistics are taken from the Mexican Social Security Institute (IMSS) databases.³As far as information on education is concerned, SEP data regarding enrollment by educational level in Mexico is used.⁴Finally, in relation to poverty, the In-work Poverty Trend Index (ITLP) constructed by the National Council for the Evaluation of Social Development Policy (CONEVAL) is studied.⁵

Estimation strategy

The econometric strategy consists of an impact evaluation analysis using the staggered adoption of the synthetic control method pursuant to the methodology of ^{Ben-Michael et al. (2021}). Said method is an extension of the original synthetic control design formulated in the seminal work of ^{Abadie and Gardeazabal (2003}) and subsequently applied in Abadie et al. (2007) and Abadie et al. (2015). The initial version of the method assumes that a set of units is observed before and after a certain intervention, where one of the units is exposed to that intervention (treatment) and the remaining units constitute a control group. The synthetic control is constructed utilizing weights that minimize the square of the differences between the characteristics of the treatment unit and the unexposed units in the period before the intervention.

Formally, in accordance with ^{Abadie and Gardeazabal (2003}), Abadie et al. (2007) and Abadie et al. (2015), if it is assumed that J + 1 units exist, where j = 1represents the treated unit, and j = 2, ..., J + 1 are the remaining untreated units. In turn, T0and T1are defined as the period before and after the intervention, with T=T0+T1. In this case, for example, the regional development trajectory of the treated unit (Y1t), p is known, but the counterfactual of what would have happened with regional development if the treated unit had not received the intervention associated with the automotive plants ( Y1tN para T0)is not known. Therefore, a timely estimate of Y1tNneeds to be calculated in order to obtain the treatment effect (α _it ), which is expressed as:

For the above, the methodology proposes that a vector of weights of the control group must be found,W=w2,…,wj+1, with 0≤wj≤1, for the untreated units j = 2, ..., J + 1, and subject to the restriction ∑j=2j+1wj=1, in such a way that the weighted average of the relevant characteristics of all the units in the control group⁶should resemble the treated unit. These characteristics are included in a vector for the treated group (X1)and a matrix for the control group (X0). The method seeks to minimize the difference in characteristics X1-X0W, where the synthetic control group, W*, is the one that minimizes these differences. Once the value of W* is obtained, it is used to estimate the counterfactual of Y1tN, which is the result of the synthetic control in the period after the intervention, i.e. ∑j=2j+1W*Yjt=1for t≥T0. Thus, the treatment effect of the unit receiving the intervention ∝1t^is :

In other words, the treatment effect is the difference between the treated unit's observed trajectory and the synthetic unit's trajectory after the intervention period, with the synthetic unit being the counterfactual of what would have happened if the treated unit had not received the treatment. A significant advantage of the synthetic control is that it allows us to empirically show the differences between the trajectory of the treatment and control groups. An excellent synthetic control should not show differences in its trajectory from the treated unit before treatment begins.

However, one of the weaknesses of this methodology is that it only allows for one treated unit to receive the intervention in a specific period of time. To date, the most common strategy for constructing synthetic controls with multiple treated units focuses on estimating separate synthetic control weights for each unit and then calculating the average of the estimates (^{Dube and Zipperer, 2015}). Nevertheless, the reliability of the results will depend on whether adequate synthetic controls can be estimated for each treated unit and whether the average is subsequently a good result for the case study (^{Ben-Michael et al., 2021}). Another option is to estimate a pooled synthetic control in which all units are averaged and differences in pretreatment are minimized as if it were a single treated unit. In this case, while a good fit for the average treated unit can be achieved, the specific fits of individual units are likely to be inadequate. Hence, the estimates will be inaccurate for individual treatment effects (Ben-Michael et al., 2021).

Pursuant to the foregoing, ^{Ben-Michael et al. (2021}) apply the methodological approach of ^{Abadie and Gardeazabal (2003}) and Abadie et al. (2007) to solve the above two problems using a method that allows for multiple treated units affected by the intervention at different times. In this case, the authors limit the error for the average treatment effect and show that this error depends on the conformation of weights on an individual and average level. They propose a partially pooled weight structure for the synthetic control method, in which the two cases described above are integrated using a hyperparameter, which indicates the relative weight of the two extreme cases in an optimization problem. Using this method, the authors enrich the synthetic control estimates by ensuring a better pretreatment fit, and by generating better individual and pooled treatment effects. The authors formally modify the synthetic control estimator making the weights the solution to a constrained optimization problem, in whichW*is calculated as follows:

Where i are the units of the control group, L are the pre-treatment periods for a particular individual i and Lis the maximum number of pre-treatment periods of the individual that takes the longest time to enter their respective treatment. The first part of equation 3 is just like the synthetic control problem. ^{Ben-Michael et al. (2021}) add differences to the treatment period (subindex l, where the weights change over time depending on the timing of the intervention) and the last term penalizes control units that are not very informative.

Based on the foregoing, the individual treatment effect ( αjk) for treated unitjat the time of treatment eventkis the difference between the potential outcome at time Tj + kfor those who receive the treatment at time Tjand the potential outcomes at time Tj + kfor units that never receive the treatment until that time (i.e., they receive it in ∞):

Given a vector 𝑁 of the optimal weights W^ijobtained from the solution of equation 3, the synthetic control estimate, YjTj+k∞, is:

Which, simply put, is the counterfactual of what would have happened to the results of a variable in question for a treated statej, at the assessed time, if the automobile plants had not established themselves in that state. Thus, the average treatment effect on the treated states ∝^jkcan be calculated with the estimated individual treatment effects.⁷

As far as the period of the study is concerned, with the exception of in-work poverty and the ITAEE, the results are estimated by applying the method to a pre-intervention period between 1997 and 2006. From 2007 to 2014, each state and/or municipality enters its individual treatment period. After this, and until 2019, the effects of treatment on each territory in the post-intervention period are estimated in order to calculate the aggregate treatment effect using the ATT. In state terms, the first federal state treated is Jalisco, with the entry of Honda in 2007. San Luis Potosí was incorporated in 2008 with the incursion of General Motors, followed by Guanajuato with Hino's investment in 2009, and then Aguascalientes and Querétaro with Nissan and VUHL in 2013 and 2014, respectively.

It is important to bear in mind that the control or donor group includes territories outside the Bajío region and the territorial states of the Bajío region, which have not yet been treated when another state in the region receives the intervention. In other words, Jalisco, the first state to receive the treatment, can include all the territories outside the Bajío region in its donor group and the other Bajío states. However, Jalisco cannot control the other territories that receive treatment later because when the Jalisco will already have been treated when the others enter treatment, the donor group may vary depending on the unit treated, the weights are not dynamic over time, i.e., the donors are fixed over time. While limiting the number of potential donors, this ensures that there is no false or unrealistic variation in the estimated counterfactuals over time due to changes in the composition of the donor group. (^{Ben-Michael et al., 2021}).

Each synthetic control estimate is based on a large set of explanatory variables pooled into five categories: economy, social conditions, education, health, security, and regressions in the dependent variables of each estimate. These variables have the same measurement period as the dependent variables and, in turn, were selected according to the availability of statistical information and theoretical postulates that affirm the possible existence of a causal relationship with the variables to be explained. Regarding the use of regressions, only those corresponding to the dependent variables in the pre-intervention period were included, with a maximum of quarterly and/or annual regressions, depending on the estimated specification. The importance of using these regressions is justified according to ^{Abadie and Gardeazabal (2003}) and Abadie et al. (2007) based on the idea that these variables improve the pretreatment adjustment of the synthetic control (the annex at the end details the variables used).

Although these variables were used to estimate a large number of models for each dimension of the assessed regional development, the final models presented in this paper are based on several different variables.

The results are those with the lowest Mean Squared Error (MSE). It should be noted that, to ensure the causality and robustness of the results, those obtained by five alternative specifications with lower MSE were compared, finding similar results both in the value and the impact trend (see the annex at the end).

Finally, statistical inference is performed using confidence intervals for the estimates of the individual treatment effects and the average ATT to validate whether the estimated effects are statistically significant. Confidence intervals are obtained by calculating standard errors of the estimates using the Jackknife resampling method, which was adapted to the staggered adoption strategy by ^{Ben-Michael et al. (2021}).

Total employment

Figure 2 shows the estimate for total employment. We can see that the fit before the intervention is good, and after that slight increases in the ATT trend are noted, which is reflected in the difference of the series in the left panel. During the first quarter, after the establishment of the new automotive companies, an average of at least 4,700 jobs had been created for each state in the Bajío region. After five quarters, average employment increased by 13,000 employees, and by quarter 10, the figure was 36,000. The effect proliferated from quarter 8, with an impact exceeding an average of 122,000 new jobs for the last quarter.

Note: The model presented is the one with the lowest MSE, including the regressions of the dependent variable as explanatory variables.

Source: Compiled by the authors.

Figure 2 Estimated staggered synthetic control for total employment 

If we compare the estimates of total employment with respect to new employment originating in the automotive industry, we can conclude that by 2019 an average of more than 122,000 total jobs had been generated in the Bajío region, of which 40,000 corresponded to the total of new direct jobs generated by the assembly plants and indirect jobs in the network of auto part suppliers and other services required by the industry. In other words, there is a differential of at least 82,000 jobs created outside the automotive sector, which indicates the presence of spillover effects of the assemblers on the manufacturing sector or the rest of the economic sectors in the Bajio region.

On aggregate, for 2019, the number of new jobs in all state economies exceeded 638,000. By comparing the two series in the left panel of Figure 2, we can calculate the difference in percentage terms of average employment with respect to its counterfactual. This shows that for 2019 there was an average increase of 15.9% in total employment. The above figure, although informative, hides the heterogeneous effects on a state level. Proportionally, the state of Querétaro registered the most significant impact with an increase of 37.4%, followed by Guanajuato, with a growth of 15.4%.On the other hand, Aguascalientes reported 13.1% new jobs, and finally, the states of Jalisco and San Luis Potosí reported rates of 6.8% and 6.4%, respectively. The results are supported by several empirical studies, which recorded similar results in terms of the effect of industry positioning on aggregate employment growth (^{Dauth, 2013}; ^{Fritsch and Mueller, 2004}; ^{Hardjoko et al., 2021}).

However, to understand the spillover effects on a territorial level, the methodology was estimated by establishing the specific municipalities that received investment in automotive assembly plants as treated.⁸The results indicate that by 2019, automotive companies had generated an average of 15,400 new jobs in the municipal economies. This represents an average increase of 20.9% in total employment in the municipalities receiving the investments. Based on the foregoing, 12.6% of the 122,000 jobs generated on a state level originated in the municipalities where the automotive plants were located and, as a result, 87.4% were jobs resulting from territorial spillover effects, which allowed the municipalities of the Bajío region to benefit from the automotive industry.

Manufacturing employment

Considering that most of the new jobs were generated outside the automotive sector, it is important to analyze whether this increase in employment occurred within the manufacturing sector or whether it was displaced to other sectors. The automotive companies' effect on manufacturing sector jobs is presented in Figure 3, which shows that the pre-treatment adjustment is good, with ATT oscillating between zero throughout the period. Once the automotive plants enter the Bajío region, the effect on employment is significant: in the first quarter, an average of 1,344 jobs were generated, which tripled three quarters later, with an increase of 4,820 employees in less than four years. The peak was reached in the last quarter when an estimated 26,500 new jobs were generated.

Pursuant to the foregoing, it is estimated that the balance of new jobs on an industrial level was less than the number of direct and indirect jobs originating from the automotive and auto parts companies.

Note: The model presented is the one with the lowest MSE, including regressions of the dependent variable as explanatory variables.

Source: Compiled by the authors.

Figure 3 Estimated staggered synthetic control for manufacturing employment 

In other words, there is a gap which suggests that, during this period, although jobs were generated in manufacturing thanks to automotive production, a significant number of jobs appear to have been lost in other industrial areas. Proportionally, the automotive industry increased manufacturing employment by an average of 14.3% for the states in the Bajío region.⁹The federal state with the most significant impact was Aguascalientes, with a growth of 20.8%. This was followed by Guanajuato, which received the largest number of assembly plants during this period, with an increase of 16.1%, followed by Querétaro and San Luis Potosí, with increases of 14.2% and 12.4%, respectively. Finally, Jalisco experienced a lower impact, with an increase of only 8.1%.

As far as municipal estimates are concerned (in the annex at the end), it is calculated that, on average for 2019, automotive companies triggered an increase of 7,245 industrial jobs in the municipalities where they were located. This increase in jobs represents an average growth of 30% in manufacturing jobs in the municipalities. Two important points can be deduced from this. First, 46.8% of the total number of jobs created in the municipalities where the automotive companies set up their businesses were in the manufacturing sector, compared with 21.7% on a state level. Second, new regional industrial jobs were concentrated to a greater extent in the municipalities receiving the investments (27.3% of the jobs), compared with new aggregate jobs, of which only 12.6% were created in these municipalities. Therefore, significant spillover effects exist in relation to the municipalities that did not receive the automotive companies.

Quarterly Indicator of State Economic Activity (ITAEE)

Figure 4 shows the estimated impact on economic activity. We can see that, after the intervention, the ATT increased considerably over time: in the first four quarters, the automotive industry managed to increase the ITAEE by an average of 4.14 pp. This impact proliferated, obtaining an increase of 10 pp in quarter 13 and remaining at that level during subsequent periods reaching a peak of 11.74 pp in quarter 20 (2019).

Note: The model presented is the one with the lowest MSE, including the IED, the unemployment rate and the regressions of the dependent variable as explanatory variables. Source: Compiled by the authors.

Figure 4 Estimated staggered synthetic control for the ITAEE 

Wages

Taking into account the aforementioned effects on the labor market and on economic activity, it is worth questioning whether this translated into improvements in wages for workers in the region. Therefore, the methodology for the average daily wage was estimated (see Figure 5). We can see that once the automotive companies are established, the ATT trend diverges from zero and grows continuously, reaching an effect greater than MXN$23 for 2019. This indicates that the automotive companies increased the daily wage of Bajío workers by an average of 11.8%. The most significant impact was observed in Queretaro, with an increase of 23.2%, which is consistent with the fact that this territory also registered the most significant expansion in total employment. The results are consistent with the evidence in international literature (^{Lee and Rodriguez-Pose, 2016}).

Note: the model presented is the one with the lowest MSE, including the unemployment rate, the ITAEE and the regressions of the dependent variable as explanatory variables.

Source: Compiled by the authors.

Figure 5 Estimated staggered synthetic control for the average daily wage (IMSS) 

In-work poverty

We then analyze whether the previous effects were enough to improve social welfare indicators, such as in-work poverty (see figure 6). After the first year of intervention, the assembly companies had managed to reduce in-work poverty by 3.0 pp on average and this effect continued time decreasing by 10.51 pp in 2019. This indicates that had it not been for the establishment of the assembly plants, the Bajío states would have had at least 10 pp more in-work poverty by 2019. This impact is significant in terms of regional development, given that thousands of people were able to escape poverty thanks to the industrial growth of the automotive sector.

Note: The model presented is the one with the lowest MSE, including employment per capita, average wage, FDI and regressions of the dependent variable as explanatory variables.

Source: Compiled by the authors.

Figure 6 Estimation of the synthetic staggered control for in-work poverty 

Had it not been for the investments in the automotive plants, by 2019 all of the Bajío states would have had in-work poverty levels above the national average. Using the counterfactual concept, a static analysis of what would have happened with the in-work poverty ranking is performed (from lowest to highest), if the assembly plants had not been established in the Bajío region (this analysis is summarized in Table 2). The greatest reduction in in-work poverty was recorded in Jalisco, which would have moved from position 6 to 25 in the ranking for lowest in-work poverty if the automotive plants had not been located in its territory. This was followed by San Luis Potosí and Guanajuato, which would have dropped six and five positions, respectively. In the case of San Luis Potosí, this state would have been among the four states with the highest proportion of people subject to in-work poverty. Finally, Aguascalientes and Querétaro would have dropped three positions, surpassing the median of the poorest states.

Figure A6 in the annex shows the behavior of in-work poverty for the five states treated and the average in-work poverty trend for the states that make up the synthetic control. The trend of poverty reduction in the treated states and the marked increase in this indicator for the synthetic control group in the post-treatment period confirms the validity of the estimates and lends veracity to the causal effect described above.

Enrollment in education

Since industrial plants may increase the demand for workers with a particular level of education, this could be expected to impact the educational enrollment records of the inhabitants of the Bajio region. The annex shows that there is no effect on higher education; however, there is an effect on high school education (see figure 7). In this respect, thanks to the automotive assembly companies, a greater number of young people attended high school, registering an increase in enrollment of 216 new students per ten thousand people in 2018, representing an average increase in enrollment of 3%.

Note: The model presented is the one with the lowest MSE, including GDP and public spending on education per capita.

Source: Compiled by the authors.

Figure 7 Estimated staggered synthetic control for high school enrollment per ten thousand people between 15 and 17 years of age 

Robustness tests

The results were subjected to various robustness tests, which permitted the assessment of sensitivity to the specification of the synthetic controls, the possible diversity of impacts across federal states, and the potential incidence of the northern and central border states. In the first case, for each dimension of regional development, at least 20 combinations of different models were estimated, including (or excluding) covariables, interactive fixed effects, and fixed effects by treated unit. Second, synthetic controls were again estimated, including three different weights when calculating ATT. Finally, all the estimates were repeated for the automotive states in the northern and central region for the donor set and excluding the states on the territorial borders of the Bajío states. In general, the results obtained are robust, the pre-treatment adjustment and the trends do not experience significant modifications given the different robustness tests, and although some variations occur in the ATT, the direction of the results does not change (results in the annex at the end).

Meanwhile, placebo tests were performed for the final synthetic controls presented in the previous section. The period of the treatment simulating an advance of the impact was changed, and the donor group entered the treated states. In both cases, no impacts are generated by the results, nor do we see trends similar to the true ones.¹⁰

5. CONCLUSIONS

This research studied the consequences of the extraordinary growth of the automotive industry in Mexico over the last two decades on regional development. To our knowledge, it is the first causal study to evaluate the impact of the automotive industry on regional development in Mexico, and it contributes to international literature on studies that have analyzed the effects of large industries on development characteristics.

On a state level, the automotive industry increased total employment and manufacturing employment by an average of 15.9% and 14.3%, respectively. The calculations indicate that for every new job created in the automotive sector, an average of 4.97 new jobs were generated in the Bajío state economies. On aggregate, more than 638,000 jobs were created in the territories of the Bajío region, of which 78.3% were created in sectors other than industrial branches. This confirms the extensive spillover effects of the automotive sector, apparently with a much more substantial impact on the service sector.

When assessing impacts within the manufacturing sector, evidence was found of possible job destruction in industrial areas other than the automotive sector, given that the absolute impact on the number of manufacturing jobs was lower than the number of jobs created in automotive and auto parts plants. This is consistent with the concerns of regional development theories in that the arrival of new industries can negatively affect other established industrial branches, generating a displacement effect on economic activity and employment. Despite this, the net employment balance continues to be positive.

The establishment of the automotive industry in the Bajío region promoted local economic development: it increased economic activity, reduced in-work poverty, and increased school enrollment on a high school level. As far as poverty is concerned, according to ^{Fowler and Kleit (2014}), its reduction seems to be explained by the fact that the majority of the new jobs are occupied by local workers, who, by obtaining a source of income, can escape the poverty line. The increase in enrollment in education is consistent with the postulates of ^{Porter (1998}) and the estimates of ^{Fu and Gabriel (2012}), who state that industrial connections continually require increased human capital, so a better supply of education and graduates in the territories would be expected.

Although different variables associated with social development have been analyzed, this paper does not study the responses of state and local governments. These responses could supplement the establishment of automotive plants in the region and explain differences in the results. This paper does not study other specific mechanisms to observe the estimated effects. These questions remain open for future research.