Local Labor Markets in U.S.-Mexican Border Cities and the Impact of Maquiladora Production. André Varella Mollick* Abigaíl Cortez Rayas Rosa A. Olivas Moncisvais

Abstract: Shift-share decompositions of employment positions created over 1990-2002 at U.S.-Mexico border cities show that El Paso employment change has been overwhelmingly attributed to national forces, while local and national forces roughly match employment creation at Brownsville. With such fast growing U.S.-Mexico border area as background, we implement time series labor demand models to two of these Texas border cities (El Paso and Brownsville) and compare movements in Mexican maquiladora production against U.S. employment and wage forces. Applying cointegration methods, we do confirm that local (purely regional) shocks are more important in Brownsville than in El Paso. Specifically, 10% increases in Juárez maquiladora production lead to increases of 1.21% in El Paso’s government employment and of 0.88% in overall employment. Similar increases in Matamoros lead to gains of 1.59% in Brownsville’s trade, transportation and utilities employment and of 1.41% in overall employment. These results support shift-share findings and are generally consistent with a more diversified industry-mix promoting higher employment creation. Keywords: border cities, employment, local labor markets, maquiladoras, Mexico. JEL Classification Numbers: F14, F16, R12.

* Corresponding Author. Department of Economics and Finance, College of Business Administration, University of Texas - Pan American, 1201 W. University Dr., Edinburg, TX 78541-2999, USA. E-mail: [email protected] Tel.: +1-956-316-7913 and fax: +1-956384-5020. Part of this research was funded by the research project on “The Economic Agenda of the Mexican Border” held by the ITESM-Campus Monterrey in Mexico. A previous version was presented at the Texas A&M International Cross-Border Workshop in Economics & Finance held in Laredo in October of 2004. The authors acknowledge, without implicating, participants of that conference, Pedro Albuquerque, Jesus Cañas, Roberto Coronado, and two anonymous referees for very useful comments on previous versions of this paper.

2 1. Introduction Maquiladoras (export assembly plants) are responsible for a substantial portion of Mexican manufacturing jobs (29% in 2000 according to Vargas (2001)) and of total Mexican exports (roughly 50% in 2001, according to Kose et al. (2004)). The degree of integration between inputs and outputs is a distinguished feature since inputs come mostly from the U.S. and exports are destined to the U.S: “Maquiladora firms are those that import nearly 100% of their inputs and then export nearly 100% of their output” (Robertson, 2003, p. 40). There are several studies on factors affecting the employment and output dynamics of Mexican maquiladoras. The papers by Fullerton and Schauer (2001), Gruben (2001), and Mollick (2003) contain estimates of maquiladora employment as function of exchange rates and U.S. economic variables. Less known, however, are studies that quantify the impact of maquiladora production on economic variables. Since most of the maquiladora establishments are located at the U.S. – Mexican border, it is important to know if their degrees of economic activity have any impact on U.S. cities. Empirical evidence on this issue is scant with perhaps two exceptions. Hanson (1996) estimates panel of annual data on the six largest border-city pairs from 1975 to 1989. He finds that a 10% increase in export manufacturing in a Mexican border city leads to between 2.4% and 4.9% increases in manufacturing employment in the neighboring U.S. border city and much smaller effects on services. On a single border city-pair, Zamora and Lecuanda (2002) document sales effects in San Diego of changes in the Mexican peso exchange rate and in Tijuana disposable income. This paper is an attempt to contribute to the second strand in this literature. As in Hanson (1996), we adopt maquiladora production at the Mexican city as the major local force to the border cities at the U.S. side. As motivation, Hanson (2002), for example, points out to the average growth rate of real value added produced by the maquiladoras at around 10% in

3 1990-2002 against the 3% average growth of real Mexican GDP during the period. This suggests strong local shocks at the U.S. border city if there is any cross-border transmission mechanism. In contrast to the strong effects on manufacturing reported by Hanson (1996), Vargas (2001, p.27) presents a convincing picture that this could be so in services. She mentions, for instance, the fact that “Texas border cities have reaped important benefits from their maquiladora neighbors. Transportation and customs services have flourished on the U.S. side of the border because of the maquiladora industry’s large trade flows through border ports of entry. These companies typically maintain distribution facilities and administrative offices on the U.S. side, stimulating the industrial real estate sector of Texas border cities. Maquiladoras also create jobs for the U.S. border in the legal, accounting and financial professions. Even the hotel, car rental and restaurant industries profit from maquiladoras because corporate personnel and other maquiladora visitors usually stay and eat on the U.S. side.” There are at least three novelties in our approach. First, in contrast to Hanson (1996)’s panel data methodology, monthly time series models are investigated ranging from OLS estimations of first-differenced models to cointegration analysis. Second, the final model comes as a natural complement to the Blanchard and Katz (1992) approach to local labor markets in the two Texas border city-pairs studied: El Paso-Ciudad Juárez and Brownsville-Matamoros. This paper thus adds to the evidence on the roles of oil price, defence, local and national shocks in explaining employment at the city level by Bhattacharya (2003), for example. On population dynamics, the Texas Workforce Commission (2003) states that, out of the five MSAs that posted the highest annual growth rates during 2002, four (Brownsville-Harlingen, El Paso, Laredo and McAllen-Edinburg-Mission) are located on the border. This would imply substantial labor demand at trade and government sectors as documented by Orrenius and

4 Berman (2002). As far as we know, there is no comparative study between city pairs with different sizes, growth rates or even labor composition. Third, the data for the two U.S. border cities in this paper show the decline of manufacturing share in total employment. From January 1990 to June 2002, the manufacturing share fell from 18% to 11% in El Paso and from 12% to 8% in Brownsville. Hitherto, the only study there is on U.S. border employment decompositions comes from GAO (2003): over 19902002 about 39% of employment creation are due to local forces, 57.4% are due to the national effect and 3.6% to the industry-mix effect.1 This suggests that local shocks are more important in Brownsville than in El Paso. However, in the shift-share literature local shocks are obtained as residual. Attributing local shocks to maquiladora production at the Mexican city, this paper provides econometric support to this conjecture. We do confirm higher local employment responses to maquiladora operations in Brownsville than in El Paso. We also find stronger local effects of maquiladora production on sectors other than manufacturing. Applying cointegration methods, a 10% increase in maquiladora production in Juárez leads to increases of 1.21% in government employment in El Paso and of 0.88% in El Paso’s overall city employment. Conversely, a 10% increase in maquiladora production in Matamoros leads to increases of 1.59% in trade employment and of 1.41% in Brownsville’s overall city employment. This paper comprises five sections. Section 2 explains the data construction and section 3 contains the empirical methodology. Section 4 presents the major results of our work and section 5 concludes with extensions for further research and policy implications. 1

Results vary across locations. At El Paso, for example, 16.59% of the total employment change is due to local forces, -10% is due to industry mix, and an overwhelming 93.45% is due to national forces. At Brownsville, 52.39% of the total employment change is due to local forces, none is due to industry mix, and 47.32% is due to national forces. See GAO (2003) and table 1 below for a summary.

5 2. Data Description According to Mexico’s Censo de Población as of 1990, Ciudad Juárez is the most populated (799,000) among Mexican cities located along the U.S.-Mexico border, followed by Tijuana (747,000). According to U.S. BEA, the largest border city, measured by metropolitan statistical area (MSA), at the U.S. side as of 1990 is San Diego (2,513,000), followed by El Paso (596,000). As San Diego is much higher than Tijuana and this city-pair is well characterized by “one-sided economic dependence” according to GAO (2003), we focus on two other city pairs: El Paso-Juárez and Brownsville-Matamoros. The former is the natural candidate to study given its population size and economic importance and the latter comes as a smaller size example of a city pair in which the Mexican city (Matamoros with 303,000) roughly matches the corresponding U.S. MSA (Brownsville with 262,000).2 Data are of monthly frequency from January of 1990 to June of 2002, obtained from different sources. Data for maquiladora value added, employed personnel, wages and consumer price index for Mexican cities (Juárez and Matamoros) are obtained from INEGI’s website (http://www.inegi.gob.mx). INEGI’s dataset comprises the Banco de Información Económica (BIE), which contains the major source of the maquiladora data used in this paper. The spot exchange rate information comes from Banxico’s website (http://www.banxico.org.mx): FX market, 48 hours interbank rate, average quotes of close prices. Data for the unemployment rate and employment figures at the MSA in U.S. border cities are obtained from the Bureau of Labor Statistics (http://data.bls.gov) of the U.S. Department of Labor. Real wage rates (average weekly earnings of production workers) included three different sectors: construction, 2

An interesting city pair to study would have been McAllen and Reynosa, which are economically interdependent, very fast growing, and medium size cities: McAllen’s MSA with 569,000 and Reynosa with 420,000, according to GAO (2003). However, value added data on maquiladora production for the city of Reynosa is incomplete and precludes the monthly time series approach below. For this reason, we focus on the medium sized city pair of Brownsville and Matamoros. A detailed descriptive study of the Tijuana-San Diego interdependence is provided by Chang-Hee (2003).

6 manufacturing, and trade, transportation, and utilities. These data are originally not seasonally adjusted. The unemployment rates in Brownsville, El Paso and Texas are shown in figure 1 and represent the number of unemployed persons as percent of labor force. The two U.S.-border cities show a decreasing trend in unemployment during the period, with some reversal in 20012002. This is consistent with U.S. border cities responding to lower economic expansion after the start of the March 2001 recession. Both city unemployment rates are much higher than the (single-digit) Texas average and Brownsville shows higher unemployment rates than El Paso throughout. [Figure 1 about here] The unemployment rate, however, does not help visualizing advances or retreats at the sector level. We thus prefer to focus on the more detailed employment statistics. Figures 2a and 2b display national employment figures in the left vertical axis against the local (El Paso and Brownsville) employment figures in the right axis. LNAT is defined as: LNAT=LUS – LTexas in order to avoid simultaneity as in Hanson (1996) and McCarthy and Steindel (1997). These figures contain national labor markets compared to local markets in the two border cities. On the national side (100,641,000 employees in January of 1990), there is stagnation in employment growth during the early 1990s and early 2000s, while El Paso and Brownsville differ in labor growth patterns. [Figures 2a and 2b about here] The decomposition of employment figures over 1990-2002 is presented in figures 3a and 3b for El Paso and Brownsville, respectively. In both cases, one confirms a reduction in the share of employment due to manufacturing (LM) activities, although more acute at El Paso, coming from 18% of employment in 1990 to slightly over 11% in 2002. The average of LM in

7 El Paso over the period is 16.5%. Brownsville also shows a reduction in LM ending the sample at only 8% and averaging 11.2%.3 Studies for other regions, such as McCarthy and Steindel (1997) for the New York metropolitan area, also refer to the “shrinkage of the regional manufacturing sector and the rapid growth of the financial and business services sectors”. In addition to manufacturing, figures 3a and 3b show the rather flat patterns of trade, transportation and utilities (LT) as well as government (LG) employment over time. These 3 sectors are chosen because of their relative importance in total employment figures. The averages of monthly figures for these 3 sectors (LM, LT and LG) add up to 56.6% of total employment in El Paso and up to 49.8% in Brownsville. Average real weekly earnings of production workers are displayed in figure 4. Both series move together during the period and share a flat common trend. [Figures 3a, 3b, and 4 about here] An interesting account of the labor turnover across sectors in Texas is provided by the Texas Workforce Commission (2003) in its assessment of the severe 2001 downturn. While educational and health services led all industry groups with an increase of 46,000 jobs during 2002, the government sector continued to create jobs, adding 30,700 positions over the year. The overall loss of 10,700 jobs in 2002 in MSAs was much smaller than the 109,000 jobs drop a year earlier and “El Paso posted the largest increase, followed by San Antonio, BrownsvilleHarlingen and McAllen-Edinburg-Mission…Out of the five MSAs that posted the highest annual growth rates during 2002, four (Brownsville-Harlingen, El Paso, Laredo and McAllenEdinburg-Mission) are located on the Texas-Mexico border. Population growth in these areas 3

Employing the location quotient technique, Cañas (2002) documents that El Paso still has a larger concentration of manufacturing jobs (16%) compared to the nation (14%), despite the fall in manufacturing employment in the 1990s. We do not know of similar studies to the Brownsville-Harlingen region. General reports refer to Brownsville as a “diverse region with a major presence in manufacturing, tourism and retailing.” (Brownsville Economic Development Council, 2004).

8 contributed to significant employment increases in Educational and Health Services, as well as in the Government sector. Brownsville-Harlingen’s annual growth rate of 4.4% was the highest for any Texas MSA” Texas Workforce Commission (2003). Table 1 reproduces calculations from GAO (2003) based on shift-share analysis.4 Of the total employment change displayed at the bottom of the table (590,800), the shift-share analysis calculates the national growth effect, the industry-mix effect, and the local effect. The national growth effect assumes the region’s employment growth matches the national trade. For example, if, during 1990-2000, U.S. non-farm employment grew by 20%, the national growth component of any region within the U.S. during this decade would be 20% of the region’s 1990 employment. Table 1 shows that from 1990 to 2002 the border counties gained 339,100 jobs due to economic trends at the national level, of which 42,800 in El Paso and 16,800 in Brownsville. The industry-mix effect is the amount of change that a region would have experienced had each of its industries grown at their industry national rates, less the national growth effect. From 1990 to 2002 the border counties gained 21,200 jobs owing to concentration of faster growing sectors there than in the nation as a whole, primarily in services (114,400). El Paso shows total L contraction due to this effect (-4,600) but none in Brownsville. The local effect is the employment change that remains after the national and industrial mix components have been accounted for. It is therefore the purely regional aspect of the region’s employment growth. Across all sectors, employment growth attributed to local conditions totals to a net addition of 230,500 jobs: 7,600 in El Paso and 18,600 in Brownsville. In addition to local government policies and incentives, this regional component should capture 4

The nine original sectors in GAO (2003)’s report match the BLS classification: construction & mining, manufacturing, transportation & public utilities, trade, wholesale trade, retail trade, FIRE (finance, insurance, and real estate), services, and government. In table 1, we combine wholesale and retail trade into “trade” and add this sector total to the “transportation & public utilities” category in order to obtain the “trade, transportation & public utilities” that appears in the econometric analysis below.

9 maquiladora activity south of the U.S.-Mexico border. The calculations in table 1 show that in Brownsville local and national forces (52% versus 47%) roughly match themselves in total employment creation, while in El Paso the local effect is much smaller than the national effect (17% versus 94%). This is also observed across sectors as 54.3% of trade, transportation and utilities and 64.1% of government jobs are due to local effects in Brownsville, against only 18.9% and 53.6% in El Paso, respectively.5 [Table 1 here] Statistics for south of the border are available upon request. Value Added (VA) in maquiladora is the amount of domestic raw material and package, wages, salaries, expenses and utilities used during the assembly process. In order to obtain real value added we deflate the original figures by the consumer price index (CPI) in each city of the sample, and express it in thousand of pesos of 1994. Employed personal is the number of workers in the maquiladora, with major growth right after the peso currency crisis. In order to have a U.S. dollar (USD) measure of real maquiladora activity, we divide the Mexican peso amount of value added in maquiladoras by the spot exchange rate and deflate by U.S. CPI prices. This is the variable created by Hanson (1996) as well. Figure 5 shows the constant pattern until the Mexican peso currency crisis of 1994-1995, therefore reducing the real USD value that affect U.S. cities in close distance with Mexico.6 If there are any transmission of shocks from maquiladora production to U.S. border cities, these should have been negative at the time of the Mexican crisis. The overall impact will depend on any compensating force at the national level. [Figure 5 about here] 5

Conversely, national forces seem to be more important in El Paso as 94.7% of trade, transportation and utilities and 54.8% of government jobs are due to national effects, against 54.3% and 42.4% in Brownsville, respectively. 6 Not reported figures show an inverse relationship between VA in pesos and unemployment rate in the U.S. border cities (U*). This would imply that, as maquiladora production strengthens, U* falls. The same holds for total wages and U* and for average wages (W/L) and U*. Robertson (2000) finds support that wages in the U.S. affect Mexican wages and concludes the border is more integrated with the U.S. than the interior.

10 3. The Empirical Models 3.1. The Blanchard and Katz (1992) Approach In a provocative paper, Blanchard and Katz (1992) search the answer to how much of the typical movement in state employment is common to all states and how much is statespecific. They proposed ran the following regression in logarithms for each state:

∆(Log Lit) = α + β∆(Log Lt) + ϖt

(1),

where: Lit is the logarithm of employment in state i at time t, Lt is the logarithm of U.S. employment at time t, and ϖt is the disturbance term. Blanchard and Katz (1992) estimate (1) by OLS under annual data from 1948 to 1990. As the average adjusted R2 turned out to be 0.66, they concluded that a large part of the year-to-year movement in the state employment is accounted for by movements in aggregate employment. They also concluded that the adjusted R2 are high (greater than 0.80) for states with a traditional manufacturing base, such as those in the Middle Atlantic and East North Central, which implies that employment in those states are very much dominated by aggregate movements. Their estimate of β in Texas, in particular, was at 0.82 and the adjusted R2 was found at 0.47. We conduct estimation of (1) for the state of Texas under monthly data as a first pass to whether fluctuations in Texas employment can be explained by national fluctuations. Our estimate of β in Texas was close to theirs at 0.875 and the adjusted R2 was much higher at 0.856. The Durbin Watson statistic remained at 2.155 and Breusch-Pagan serial correlation LM test rejected the null of no serial correlation. This suggests (1) could be better specified in order to take into account features of local labor markets. Introducing an additional lag to the right

11 hand side of (1) induces a statistically significant coefficient of –0.125 but no changes in the misspecification of the equation, while further lags did not turn out to be statistically significant. Table 2 reports employment at the two pairs of cities as the dependent variables and the employment of the state of Texas (LTX) as the independent variable in a county-level version of (1):

∆(Log Lit) = α + β1∆(Log LTXt) + β2∆(Log LTXt-1) + νt

(2)

We find now estimates of β1 to be 0.517 without lags and 0.552 with 1 lag of LTX for El Paso and much lower estimates for Brownsville of β1 to be 0.168 without lags and 0.184 with 1 lag. Wald tests on the β1 coefficients reject the null that β1 = 1 in both specifications for both cities. Compared to the evidence on (1) with β = 0.875, the transmission of state-level shocks to local (county) labor markets seems to be lower. The remaining columns in table 2 estimate a similar specification, now modified to the sector level as the dependent variable (j = M for manufacturing, T for trade, transportation, and utilities, and G for government) and for Li as the city (i = El Paso or Brownsville) employment level as independent variable:

∆(Log Lijt) = α + β1∆(Log Lit) + β2∆(Log Lit-1) + ηt

(3)

Results of running (3) vary depending on the sector but do have some similarity after all across cities. Table 2 shows that, for El Paso, large and close to 1 coefficients are found for LM, greater than 1 values are obtained for LT and much lower 0.30 is estimated for LG. For

12 Brownsville, the β1 coefficients vary between 0.88 and 0.97 for LM, 0.77 and 0.79 for LT, and –0.79 and –1.06 for LG. These would suggest that city employment at the government sector is less responsive to city overall labor growth in both cities. In Brownsville, the response is even negative: as city overall employment grows, city employment at the government sector falls. This may be because other sectors create more employment in an economic expansion. Conversely, in a slump, government sector job growth expands. Note, however, that the evidence of serial correlation is widespread in the estimations of (3).

3.2. The Benchmark Model Given the apparent misspecification in equations (1) to (3) and the different patterns of employment growth in figures 2a and 2b, this paper follows Hanson (1996) who presents a reduced-form version of labor demand and supply equilibrium.7 We postulate that the employment (Lijt) figures of U.S. border cities (i = El Paso or Brownsville) across sector of activity (j = manufacturing, trade, or government) can be explained by three factors, as follows:

Lijt = F (WAijt, VASit, LNATit,) + εijt

(4),

where: WA represents the alternative wage. As the wage rate at each MSA is not available, the national average by sector is used as proxy. It is measured by the average real weekly earnings 7

A more elaborate alternative to justify a reduced form such as (4) would be to suppose quadratic adjustment costs and solve for profit maximization in order to obtain equilibrium labor demand (L*) as in Hamermesh (1993). Assuming static expectations about wages and prices, the optimal path of employment is described by dLt = γ[L* Lt]. If one replaces L* by a linear function that relates L* to X (a vector composed of variables typically present in labor demand studies), one has: ∆Lt = γ’[G(Xt) - Lt-1], of which a representative empirical equation is: Lt = αLt-1 + βXt + εt, where α and β are parameters to estimate. This is the equation to be estimated below in first-differenced form by OLS and in levels by cointegration methods. Note that the employment equations in Bernanke (1986) and Brunello (1989) can be derived from first principles and yield Lt = α0 + α1yt - α2wt + α3t + εt, where: y is output, w is the real wage, and t (the time trend) is a proxy for the capital stock.

13 paid to the trade sector in case of labor growth in manufacturing (LM) or in government (LG). It is measured alternatively as real average weekly earnings paid to the manufacturing sector in case of labor growth in trade, transportation and utility services (LT) used as dependent variable.8 VAS represents in real USD the value added generated by the maquiladoras in the Mexican city of either Ciudad Juarez in case of El Paso or Matamoros in case of BrownsvilleHarlingen. The term εijt is the typical white-noise error term. As the alternative wage goes up in a given sector, employment in the other sector goes down if workers leave the latter in order to take advantage of higher wages elsewhere. A negative coefficient is thus expected on WA. As value added in Mexican maquiladoras rises, we expect economic conditions to improve along the U.S. border and thus the employment at the Metropolitan Statistic Area (MSA) in the U.S. city to grow: a positive coefficient on VAS is expected. Finally, as LNAT rises, we expect more employment growth in each sector. Note that in this framework exchange rate effects are already captured in VAS. Institutional changes, on the other hand, could be captured by NAFTA issues and a U.S.-Mexico integration dummy variable.9 Assuming logarithms, the equation to be estimated becomes:

Log Lijt = β0 + β1 Log WAijt + β2 Log VASit + β3 Log LNATit + εijt

8

(5),

In case of LG, we also used the wage of manufacturing as the relevant alternative wage. The average weekly earnings paid to production workers in the construction sector are also used as the alternative wage to all sectors (LM, LT, and LG). The results are very similar in both cases, which may reflect the fact that the wage series share common trends at the national level. Recall figure 4 above. 9 The NAFTA dummy variable defined as 0 for the years 1990 to 1993 and as 1 for the years 1994 to 2001, has been included in versions of (4) and altered some of the estimated coefficients. As it did not improve serial correlation, we prefer to benchmark our model without the NAFTA dummy variable. See Gruben (2001) on this approach and Kose et al. (2004) for more general evidence of NAFTA effects.

14 where β0 is the intercept term. Eq. (5) can be differenced to yield dynamic labor demand functions:

∆(Log Lijt) = β1∆(Log WAijt) + β2 ∆(Log VASit) + β3∆(Log LNATijt) + νt

(6),

where: νt = εijt - εijt-1. Given that (5) is usually plagued by serial correlation problems and by unit roots in the data, (6) is preferred from an econometric standpoint and can be efficiently estimated by OLS. The other route pursued below is the cointegration procedure that estimates (5) by maximum likelihood within the system framework of Johansen (1988).10

4. Results 4.1. Preliminaries and OLS Estimations For correlation coefficients calculation and OLS estimations, all series are in firstdifferences. Correlation does not appear to be a problem among right hand side variables, except for real manufacturing wages and national growth employment correlation coefficient (0.61) and for national employment and maquiladora production (0.57) for the El Paso-Juárez city pair. Underlying the more general model is the assumption that national wage rate moves together with the alternative wage at the city. Unit root tests document non-stationarity for the series in levels and stationarity in first differences.11 We thus proceed assuming the data are difference-stationary.

10

A natural generalization of (6) is the vector autoregressions (VAR) framework, in which each series is potentially endogenous. In vector form, Zt = AZt-i + υt, where Z is the vector of variables that comprises (6), A is the matrix of coefficients and υt is the vector of structural errors. 11 We employ the methodologies of Augmented Dickey Fuller (ADF), Kwiatkowski-Phillips-Schmidt-Shin (KPSS), Elliott-Rothenberg-Stock (DF-GLS), and Phillips-Perron (PP). The results are not overwhelmingly supportive of integrated of order one processes (not all 4 tests support unit roots in the data) but are certainly not supportive of

15 Table 3 reports the results from OLS estimations for El Paso and Brownsville as a first pass on the model presented in section 3. The top part of the table shows the results of estimations for El Paso. Whenever LNAT is included in the regression, it has a positive effect as expected on trade employment (0.589 for 6a and 0.902 for 6c) and on government services (0.960 for 6a and 0.770 for 6c), but weaker in manufacturing (0.371 only for 6c). For Brownsville, the β3 coefficient is always statistically significant, varying from 0.575 for 6a and 0.787 for 6c in trade, to 1.451 for 6a and 1.035 for 6c in government services, and to 0.641 for 6a and 0.775 for 6c in manufacturing. These results imply that national forces seem to exert a positive effect across all sectors, with the exception of manufacturing in El Paso. The effects of alternative wages differ across equations. In both cities, they vary from positive and statistically coefficients in LT in both city pairs to negative effects in LG equations. The β2 coefficients associated with maquiladora real value added are always statistically significant in the equations without LNAT. Their estimated values vary from 0.039 in LT to 0.055 in LG and to 0.085 in LM for El Paso. In this sense, this confirms the general finding of Hanson (1996) who obtained larger effects on manufacturing than in other sectors for a panel of six border city-pairs. In Brownsville, however, the results are reversed: from the lowest 0.021 in LM to 0.029 in LT and to 0.049 in LG. Note that there appears to be serial correlation problems for the trade sector in El Paso and for the trade and government sectors in Brownsville. In these cases, differencing does not remove specification problems. In contrast, for manufacturing employment, positive effects of changes in VAS are found as well as changes in LNAT, with the former stronger for El Paso than for Brownsville. stationary processes. Brown et al. (1990) examine cointegration between local and national employment within the same sector. They find that industrial output individual states is not cointegrated with national output in the same sector and that national and regional earnings do not cointegrate as well. These appear to contradict Altonji and Ham (1990) for Canada who document the bulk of variability in regional output due to national shocks. Coulson (1999) refers to Brown et al. (1990) and estimates sector MSA employment adopting the shift-share model.

16 4. 2. Estimates of Cointegrating Vectors This section checks whether the previous results are robust to the assumption of nonstationary series. Since the data are not seasonally adjusted, we include 11 seasonal dummies in the monthly estimates of the VAR, which turn out to be statistically significant in most cases. For the rationale on the seasonality terms, see the patterns in some employment figures, particularly in figure 2 for government employment. The estimated VARs are stable (stationary) if all roots have modulus <1, a condition observed in all models and specifications. On the lag-length selection procedure, we start with a maximum of 12 lags and check the five information criteria (LR, FPE, AIC, SC, and HQ). We next employ the serial correlation LM test of Breusch-Pagan and confirm the VARs are free of serial correlation. We report in tables 4 and 5 the results of these tests at lag orders 4 and 8, in which we always see that the systems are well specified. It almost never rejects the null of no serial correlation at h lags. The column at the right of tables 4 and 5 includes the information criteria as justification for the chosen lag length. In general, 3 lags were adopted for overall employment at the city level and 2, 3, 4 or 6 lags were used for the sector level estimates. At the top of table 4, the null of no cointegrating vectors for El Paso is rejected at the 1% level: 35.24 for the maximal eigenvalue test and 71.01 for the trace test. The latter also suggests there is more than one vector and the former indicates only one cointegration relationship. The full model, normalized for El Paso aggregate employment, yields a negative (-2.013) coefficient for wages, positive coefficient for maquiladora production (0.088), and positive coefficient of national labor effects (1.309). The signs of these estimates are all theoretically as expected and statistically significant at standard significance levels. A negative coefficient of real wages, in particular, is consistent with the following labor demand

17 interpretation: as wage rates at the retail trade sector go up nationally, city overall employment falls at the city level. More important for our hypothesis testing on the transmission of shocks across the border is the inference on VAS. As maquiladora production rises in real terms in Ciudad Juárez, the level of employment in El Paso also rises. Thus, as VAS fell with the currency crisis during 1995, employment in U.S. border cities fell. The impact is estimated at 0.088% in response to a 1% increment in maquiladora production. How does the model perform at the sector level decomposition in El Paso? It turns out that very differently. For instance, at the manufacturing sector, wages and maquila production are not statistically significant but national employment fluctuations are. However, the elasticity of LM to LNAT is estimated negative at -1.712, not a very robust figure given the standard error but certainly not negligible. This negative coefficient would be consistent with increases in national employment leading to less manufacturing growth, a phenomenon that links the “deindustrialization” observed in El Paso over the period to national factors. For the trade, transportations and utility sector in El Paso, manufacturing wages have a negative impact (-2.121) on labor demand, while national labor fluctuations have positive effects (1.180). These signs are as expected theoretically and statistically significant at standard levels. A negative coefficient of real wages implies that as wage rates go up at the manufacturing sector nationally, employment at the retail sector is reduced at the city level. Contrary to the impact at the overall city level, Mexican maquiladora production does not have a substantial impact on city trade employment since the standard errors are large (0.034). Maquiladora production certainly does have an impact, however, on the El Paso government sector employment. As real maquiladora production rises in Juárez, the level of government employment in El Paso also rises. The impact on LG is larger than the overall impact: 0.12% in response to a 1% increment in maquiladora production. Government

18 employment does not respond to wage growth in the trade sector but national employment fluctuations have positive effects at 1.189. Given that the impact of maquiladora production on total El Paso employment is positive (0.088), the impact is more visible on the government sector (0.121) than on manufacturing (-0.028) or on trade (0.028). This contrasts with previous findings by Hanson (1996) who attributed more transmission within manufacturing.12 For Brownsville, at the top of table 5 the null of no cointegrating vectors at the city level is rejected also at the 1% level: 46.30 for the maximal eigenvalue test and 70.16 for the trace test. As the subsequent null hypotheses are not rejected, there is only one cointegration relationship. The full model estimated for aggregate employment in Brownsville yields much stronger coefficients that those reported for El Paso. We find now negative (-4.129) coefficient for wages, positive coefficient for maquiladora production (0.141), and positive coefficient of national labor effects (3.016). These signs are all theoretically as expected and statistically significant at standard significance levels. As maquiladora production rises 1% in real terms in Matamoros, the level of employment in Brownsville rises by 0.141%. In Brownsville, at the sector level, the model again performs very differently. At the manufacturing sector, only wages are statistically significant at –2.449. For the trade, transportations and utility sector, manufacturing wages have a negative impact (-13.320) on labor demand, while national labor fluctuations have positive effects (3.913). And now we have the strongest effects of maquiladora production on the Brownsville trade employment sector: 0.159% in response to a 1% increment in maquiladora production. For the government

12

Hanson (1996) estimates a panel of annual data on the six largest border-city pairs from 1975 to 1989. He finds that a 10% increase in export manufacturing in a Mexican border city leads to between 2.4 and 4.9% increases in manufacturing employment in the neighboring U.S. border city. A 10% increase in Mexican export production leads to an increase in employment by 1.7% to 2.8% in the U.S. border transport industry and 1.4% to 2.4% in the U.S. border wholesale-trade industry. In services, the industry less affected, experiences an employment rise of 1.3% to 1.6%. In his model, industry dummy variables interact with the logarithm of maquiladora value added.

19 sector, wages and maquiladora production do not have significant effects while national labor fluctuations contribute positively to an enlargement of the government sector by 1.550. Since unemployment rates are traditionally higher in border cities than in Texas and in the U.S., policy implications are worth exploring. We have seen that a recovery in the U.S. economy will translate itself into expansions at the trade, transportation and utilities sector as well as in the government sector in both cities. The lack of local manufacturing employment effect in both cities suggests that an export boom in Mexico would not add to employment build-up at manufacturing firms at the U.S. side of the border. Very likely a boom in Mexican exports would generate demand to the trade, transportation and utilities as well as government at the U.S. cities as conjectured by Vargas (2001). In order to benefit from an export boom in Mexican maquiladoras, U.S. cities should encourage a large and diversified industry-mix. If so, a lacklustre U.S. economy could be compensated at the border communities with positive spillovers from Mexican maquiladoras, found to be large at the government sector in El Paso (+0.121) and at the trade, transportation and utilities sector in Brownsville (+0.159). Orrenius and Berman (2002) have noted that the wholesale and retail trade sectors are dependent on the inflow of Mexican shoppers, while certain government positions (U.S. border patrol and airport security) are more demanded when there is tightening at border crossings, such as right after the September 2001 terrorist attacks. We would expect LT to suffer from the 1994-1995 peso depreciations and LG to be affected negatively by the aftermath of the September 2001 attacks. During the period covered in this analysis, the population growth at the border has increased substantially over Texas growth rates. This likely pushes up demand for both government and trade services but not for manufacturing, which rather responds to chains of industrial clusters or location-specific initiatives.

20 5. Concluding Remarks This paper starts with the Blanchard and Katz (1992) approach to local labor markets and implements a time series version of the labor demand model in Hanson (1996) to two border city pairs: El Paso-Ciudad Juarez and Brownsville-Matamoros. Shift-share decompositions taken from GAO (2003) on employment positions created over 1990-2002 at U.S.-Mexico border cities show that at El Paso total employment change has been overwhelmingly attributed to national forces, while at Brownsville local and national forces roughly match total employment creation. Applying cointegration methods, we do confirm that local (purely regional) shocks are stronger in Brownsville than in El Paso. Specifically, 10% increases in Juárez maquiladora production lead to increases of 1.21% in El Paso’s government employment and of 0.88% in overall employment. Similar increases in Matamoros maquiladora production lead to increases of 1.59% in Brownsville’s trade, transportation and utilities employment and of 1.41% in overall employment. Contrary to the shift-share studies, however, we quantify local effects instead of obtaining them as residuals. Among the policy implications, the results suggest that a boom in Mexican exports would generate strong demand at the trade, transportation and utilities as well as government sectors at the two U.S. border cities. This paper thus sheds light on the large impact maquiladora production should have on sectors other than manufacturing, as put forward by Vargas (2001). Our findings are also broadly consistent with a more diversified industry-mix promoting higher employment creation, as documented by Orrenius and Berman (2002). Extensions include exploring local sector responses to specific shocks: the peso crisis of 1994-1995 on the trade sector and the terrorist attacks of 2001 on government positions after stronger border enforcement. VAR models that enlighten the transmission of shocks along the lines of Dávila et al. (2002) and Bhattacharya (2003) - are left for further work.

21 References Altonji J, Ham J (1990) Variation in Employment Growth in Canada: The Role of External, National, Regional, and Industrial Factors. Journal of Labor Economics 8 (1), Part 2: S196-S236. Bernanke B (1986) Employment, Hours, and Earnings in the Depression: An Analysis of Eight Manufacturing Industries. American Economic Review 76: 82-109. Bhattacharya R (2003) Sources of Variation in Regional Economies. The Annals of Regional Science 37: 291-302. Blanchard O, Katz L (1992) Regional Evolutions. Brooking Papers on Economic Activity 1: 161. Brown S, Coulson E, Engle R (1990) Noncointegration Econometric Evaluation of Models of Regional Shift and Share. NBER Working Paper 3291. Browsville Economic Development Council (2004) Major Industries. Brownsville Overview. Brunello G (1989) The Employment Effects of Shorter Working Hours: An Application to Japanese Data. Economica, Vol. 56, No. 224: 473-486. Cañas, J (2002) A Decade of Change: El Paso’s Economic Transition of the 1990s. Federal Reserve Bank of Dallas, El Paso Business Frontier 1. Chang-Hee, C (2003) Tijuana-San Diego: Globalization and the Transborder Metropolis. The Annals of Regional Science 37: 463-477. Coulson N (1999) Sectoral Sources of Metropolitan Growth. Regional Science & Urban Economics 29: 723-743. Dávila, A, Pagán J, Soydemir G (2002) The Short-term and Long-term Deterrence Effects of INS Border and Interior Enforcement on Undocumented Immigration. Journal of Economic Behavior & Organization 49: 459-472. Fullerton T, Schauer D (2001) Short-Run Maquiladora Employment Dynamics. International Advances in Economic Research 7: 471-478. GAO (2003) Mexico’s Maquiladora Decline Affects U.S. Mexico Border Communities and Trade. United States General Accounting Office. Gruben W (2001) Was NAFTA Behind Mexico’s High Maquiladora Growth?. Economic and Financial Review, FRB of Dallas, Third Quarter: 11-21. Hamermesh D (1993) Labor Demand. Princeton Univ. Press. Princeton, NJ. Hanson, G (2002) The Role of Maquiladoras in Mexico’s Export Boom. University of California-San Diego, manuscript. Hanson, G (1996) U.S. – Mexico Integration and Regional Economies: Evidence from BorderCity Pairs. (January). NBER Working Paper 5425. Johansen S (1988) Statistical Analysis of Cointegration Vectors. Journal of Economic Dynamics and Control 12: 231-254. Kose M, Meredith G, Towe C (2004) How has NAFTA Affected the Mexican Economy? Review and Evidence. (April). IMF Working Paper.` McCarthy J, Steindel C (1997) National and Regional Factors in the New York Metropolitan Economy. FRBNY Economic Policy Review, February: 5-19. Mollick A (2003) Employment Determination at Mexican Maquiladoras: Does Location Matter? Journal of Borderlands Studies 18 (2): 45-67. Orrenius, P, Berman, A (2002) Growth on the Border or Bordering on Growth. Federal Reserve Bank of Dallas, Southwest Economy 3, May/June: 1-8. Robertson R (2003) Exchange Rates and Relative Wages: Evidence from Mexico. The North

22 American Journal of Economics and Finance 14: 25-48. Robertson R (2000) Wage Shocks and North-American Labour-Market Integration. American Economic Review 90 (4): 742-764. Texas Workforce Commission (2003) Texas Labor Market Review. Labor Market Information Department, April. Vargas, L (2001) Maquiladoras: Impact on Texas Border Cities. Federal Reserve Bank of Dallas, June: 25-29. Zamora F, Lecuanda J M (2002) Interdependencia Comercial de Tijuana y San Diego. Comercio Exterior 52 (8): 680-686.

23 Figure 1. Unemployment Rates (% of Labor Force) in Texas (UTX), El Paso (UELPASO), and Brownsville (UB)

18 16 14 12 10 8 6 4 2 0 Jan-90

Jan-91

Jan-92

Jan-93

Jan-94

Jan-95 UTX

Jan-96

Jan-97

UELPASO

Jan-98

Jan-99 UB

Jan-00

Jan-01

Jan-02

24 Figure 2a. Local (El Paso) and National Employment Growth National (left axis) and El Paso (right axis) Labor Growth in Number of Employees 140000000

280000

120000000

270000 260000

100000000

250000 80000000 240000 60000000 230000 40000000

220000

20000000

210000

200000 0 Jan-90 Jan-91 Jan-92 Jan-93 Jan-94 Jan-95 Jan-96 Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02

LNAT

LElPaso

Figure 2b. Local (Brownsville) and National Employment Growth National (left axis) and Brownsville (right axis) Labor Growth in Number of Employees 140000000

140000

120000000

120000

100000000

100000

80000000

80000

60000000

60000

40000000

40000

20000000

20000

0 Jan-90 Jan-91 Jan-92 Jan-93 Jan-94 Jan-95 Jan-96

LNAT

0 Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02

LBrowns

25 Figure 3a. Decomposition of Employment in El Paso, TX El Paso Employment (L) Decompositions: LM(manufacturing), LT(trade), and LG (government) 0.24 0.22 0.2 0.18 0.16 0.14 0.12 0.1 Jan-90 Jan-91 Jan-92 Jan-93 Jan-94 Jan-95 Jan-96 Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02

LMshare

LTshare

LGshare

Figure 3b. Decomposition of Employment in Brownsville, TX. Brownsville Employment (L) Decompositions: LM (manufacturing), LT (trade), LG (government) 0.23 0.21 0.19 0.17 0.15 0.13 0.11 0.09 0.07 0.05 Jan-90

Jan-91

Jan-92

Jan-93

Jan-94

Jan-95

LMshare

Jan-96

Jan-97

LTshare

Jan-98

Jan-99

Jan-00

LGshare

Jan-01

Jan-02

26

Figure 4. Real National Wage Rates in Manufacturing and Trade Average Real Weekly Earnings of Production Workers in Manufacturing (WMP) and Trade (WTP) 380 360 340 320 300 280 260 240 220 200 Jan-90

Jan-91

Jan-92

Jan-93

Jan-94

Jan-95

Jan-96

WMP

Jan-97

Jan-98

WTP

Jan-99

Jan-00

Jan-01

Jan-02

27 Figure 5. Maquiladora Production (Real Value Added in USD) in Ciudad Juárez and Matamoros Value Added of Mexican Maquiladora Production in Cd. Juárez and Matamoros (in real U.S. dollars) 160000 140000 120000 100000 80000 60000 40000 20000 0 Jan-90

Jan-91

Jan-92

Jan-93

Jan-94

Jan-95

Jan-96

VASElPaso

Jan-97

Jan-98

Jan-99

VASBrowns

Jan-00

Jan-01

Jan-02

28 Table 1. Components of Employment Changes by Sectors in U.S.-Mexico MSAs, 19902002 from GAO (2003).

Local Effect U.S.-Mexico Border El Paso Brownsville IndustrialMix Effect U.S.-Mexico Border El Paso Brownsville National Effect U.S.-Mexico Border El Paso Brownsville

Total L

% of total ∆L

Manufacturing

% of total ∆L

Trade, Transp. & Utilities

% of total ∆L

Government

% of total ∆L

230.5

39.0

37.5

487.0

52.5

39.2

67.4

52.6

7.6 18.6

16.6 52.4

-5.1 -0.2

-52.0 -11.8

2.5 5.1

18.9 54.3

9.0 5.9

53.6 64.1

21.2

4.0

-69.8

-906.5

-14.4

-10.7

-10.5

-8.2

-4.6 0

-10.0 0.0

-12.8 -3.8

130.6 223.5

-1.9 -0.7

-14.4 -7.4

-1.4 -0.6

-8.3 -6.5

339.1

57.4

40.0

519.5

95.9

71.5

71.2

55.6

42.8 16.8

93.5 47.3

8.1 2.3

-82.7 -135.3

12.5 5.1

94.7 54.3

9.2 3.9

54.8 42.4

Total Employment Change U.S.-Mexico 590.8 Border El Paso 45.8 Brownsville 35.5

7.7

134.1

128.1

-9.8 -1.7

13.2 9.4

16.8 9.2

Source: Figures are in thousands of employees for total employment figures and the associated percentages of total employment change, calculated with respect to the figures at the bottom of the table. The entries of employment creation are from GAO (2003) based on shift-share analysis. The original sectors in GAO’s report match the BLS classification and are nine: construction & mining, manufacturing, transportation & public utilities, trade, wholesale trade, retail trade, FIRE (finance, insurance, and real estate), services, and government. We combine wholesale and retail trade into “trade” and add this “trade” total to the “transportation & public utilities” category in order to obtain the “trade, transportation & public utilities” that is going to appear in the econometric analysis below. We omit the remaining sectors since they are not relevant to our analysis.

29 Table 2. OLS Estimations of Blanchard and Katz (1992) Equations ∆(Log Lit) = α + β1∆(Log LTXit) + β2∆(Log LTXit-1) + νt

(2),

∆(Log Ljt) = α + β1∆(Log Lit) + β2∆(Log Lit-1) + ϖt

(3)

El Paso-Juárez City Pair

α β1

Eq. (2)

Eq. (2)

Eq. (3)

Eq. (3)

Eq. (3)

Eq. (3)

Eq. (3)

Eq. (3)

∆log(L) -0.0001 (0.0006) 0.517*** (0.072)

∆log(L) -0.0005 (0.0006) 0.552*** (0.071) 0.221*** (0.071)

∆log(LM) -0.003** (0.001) 0.886*** (0.166)

∆log(LM) -0.003** (0.001) 0.875*** (0.162) 0.494*** (0.162)

∆log(LT) -0.0001 (0.001) 1.140*** (0.121)

∆log(LT) -0.0003 (0.001) 1.133*** (0.121) 0.181 (0.121)

∆log(LG) 0.001 (0.002) 0.263 (0.192)

∆log(LG) 0.0021 (0.0014) 0.301* (0.179) -0.875*** (0.179)

0.255 1.919 44.99*** [0.00]

0.300 1.964 17.59*** [0.00]

0.156 2.207 0.47 [0.50]

0.204 2.348 0.59 [0.44]

0.373 1.753 1.39 [0.25]

0.378 1.784 1.21 [0.27]

0.006 1.680 8.08** [0.02]

0.141 1.519 15.26** [0.0001]

0.850 [0.654]

0.478 [0.787]

15.21*** [0.0005]

22.29*** [0.00]

16.33*** [0.00]

12.75*** [0.00]

14.78*** [0.00]

16.10*** [0.00]

β2 Adj. R2 DW Wald F-test: β1=1 LM-test

Brownsville-Matamoros City Pair

α β1

Eq. (2)

Eq. (2)

Eq. (3)

Eq. (3)

Eq. (3)

Eq. (3)

Eq. (3)

Eq. (3)

∆log(L) 0.002*** (0.0007) 0.168*** (0.081)

∆log(L) 0.002*** (0.0007) 0.184*** (0.082) 0.108 (0.082)

∆log(LM) -0.002 (0.002) 0.876*** (0.192)

∆log(LM) -0.003** (0.0016) 0.973*** (0.196) 0.415** (0.197)

∆log(LT) 0.0002 (0.0014) 0.769*** (0.169)

∆log(LT) 0.000 (0.0015) 0.790*** (0.174) 0.087 (0.175)

∆log(LG) 0.004** (0.002) -0.785*** (0.279)

∆log(LG) 0.008*** (0.002) -1.059*** (0.272) -1.172*** (0.273)

0.022 2.450 105.16*** [0.00]

0.027 2.440 98.79*** [0.00]

0.118 2.201 0.42 [0.52]

0.138 2.29 0.019 [0.89]

0.118 1.938 1.88 [0.17]

0.114 1.936 1.46 [0.23]

0.045 1.714 41.07*** [0.00]

0.146 1.764 57.37*** [0.00]

β2 Adj. R2 DW Wald F-test: β1=1 LM-test

8.209** 7.823** 3.551 5.40** 6.657*** 5.91* 14.52*** 17.21*** [0.017] [0.02] [0.169] [0.07] [0.036] [0.052] [0.001] [0.0002] Notes: The regressions employ the Newey-West heteroscedastic-autocorrelation consistent (HAC) matrix, with lag truncation = 4 and standard errors in parentheses. The symbols * , ** , ***, indicates statistical significance at 10, 5 and 1 percent level, respectively. The sample size is either 150 or 149 observations after adjusting end points.

30 Table 3. OLS Estimations of Employment Equations in El Paso and Brownsville ∆(Log Lijt) = β1∆(Log WAitj) + β2∆(Log VASitj) +β3∆(Log LNATt) + νt

(6a),

∆(Log Lijt) = β1∆(Log WAitj) + β2∆(Log VASitj) + νt

(6b),

∆(Log Lijt) = β2 ∆(Log VASit) + β3∆(Log LNATijt) + νt

(6c),

El Paso-Juárez City Pair Dep. Variable/

Model

β1 β2 β3

LM 6a 0.138 (0.126) 0.070*** (0.018) 0.296 (0.198)

LM 6b 0.203* (0.118) 0.085*** (0.015)

LM 6c

0.068*** (0.018) 0.371** (0.186)

0.190 0.196 Adj. 0.197 2 R 1.763 1.687 DW 1.720 0.242 0.000 1.128 LM[1.00] [0.569] test [0.886] Brownsville-Matamoros City Pair

LT 6a 0.355*** (0.071) 0.019 (0.012) 0.589*** (0.138)

LT 6b 0.490*** (0.067) 0.039*** (0.012)

LT 6c

LG 6b -0.135 (0.136) 0.055*** (0.017)

LG 6c

0.034** (0.013) 0.902*** (0.133)

LG 6a -0.346** (0.136) 0.005 (0.020) 0.960*** (0.215)

0.499

0.440

0.416

0.162

0.054

0.131

1.776 22.345*** [0.00]

1.811 15.811*** [0.00]

1.541 30.11*** [0.00]

1.920 0.171 [0.918]

1.831 3.270 [0.195]

1.726 3.365 [0.186]

LG 6b -0.638*** (0.185) 0.049*** (0.015)

LG 6c

0.010 (0.020) 0.770*** (0.205)

Dep. Variable/

Model

β1 β2 β3 Adj. R2 DW LMtest

LM 6a 0.295* (0.153) 0.004 (0.013) 0.641** (0.278)

LM 6b 0.419** (0.180) 0.021** (0.008)

LM 6c

LT 6b 0.391*** (0.097) 0.029*** (0.007)

LT 6c

0.004 (0.014) 0.775** (0.315)

LT 6a 0.256*** (0.095) 0.019** (0.007) 0.575*** (0.158)

0.027*** (0.007) 0.787*** (0.148)

LG 6a -0.919** (0.181) 0.009 (0.011) 1.451*** (0.285)

0.139

0.095

0.120

0.321

0.280

0.294

0.224

0.098

0.106

2.051 1.315 [0.518]

2.048 2.850 [0.24]

2.060 1.309 [0.520]

1.904 15.870*** [0.00]

1.918 11.360*** [0.003]

1.800 15.919*** [0.00]

1.670 12.394*** [0.002]

1.718 14.986*** [0.0006]

1.463 21.467*** [0.00]

0.008 (0.014) 1.035*** (0.286)

Notes: The regressions employ the Newey-West heteroscedastic-autocorrelation consistent (HAC) matrix, with lag truncation = 4 and standard error in parentheses. The symbols * , ** , ***, indicates statistical significance at 10, 5 and 1 percent level, respectively. The sample size is either 150 or 149 observations after adjusting end points.

31 Table 4. Cointegration Tests in El Paso of Full Model in Logarithms: [L, wages, VAS, LNAT] Null Hyp. VAR Lag Trace 5% 5% Max. Model Specification and Critical on Coint. Length / Eigenv. Critical Test Estimation of Johansen Value Vectors LM Test Value Test Cointegration Vectors FULL MODEL – CITY 35.24** 27.07 71.01** 47.21 3 by FPE, No C.V.s Log(L)= -2.013 log(WTP) 19.01 20.97 35.77** 29.68 At most 1 AIC, HQ (0.401) 13.66 14.07 16.77* 15.41 At most 2 3.11 3.76 3.11 3.76 At most 3 19.67 17.50 + 0.088 log(VAS) [0.24] [0.35] (0.029) +1.309 log(LNAT) (0.164) FULL MODEL - MANUFACTURING

Log(LM)= -1.897 log(WTP) (2.072) -0.028 log(VAS) (0.146) -1.712 log(LNAT) (0.892)

29.74* 22.57* 16.45* 6.84**

27.07 20.97 14.07 3.76

75.61** 45.86** 23.29* 6.84**

47.21 29.68 15.41 3.76

No C.V.s At most 1 At most 2 At most 3

3 by LR, FPE, AIC, HQ

30.81* 18.46 4.00 1.77

27.07 20.97 14.07 3.76

55.04** 24.24 5.78 1.77

47.21 29.68 15.41 3.76

No C.V.s At most 1 At most 2 At most 3

2 by LR, FPE, AIC, HQ

28.38* 14.93 7.49 1.50

27.07 20.97 14.07 3.76

52.29* 23.92 8.99 1.50

47.21 29.68 15.41 3.76

No C.V.s At most 1 At most 2 At most 3

4 by FPE, AIC,

15.70 18.36 [0.47] [0.30]

FULL MODEL – TRADE

Log(LT)= -2.121 log(WMP) (0.477) + 0.028 log(VAS) (0.034) + 1.180 log(LNAT) (0.141)

8.75 16.07 [0.92] [0.45]

FULL MODEL - GOVERNMENT

Log(LG)= +0.885 log(WTP) (0.719) +0.121 log(VAS) (0.051) +1.189 log(LNAT) (0.292)

14.36 12.66 [0.57] [0.70]

Notes: Intercept and trend are included in cointegrating equation and test VAR.The data in levels have linear trends but the cointegrating equation has only intercepts. The symbols ** indicates significance (rejection of the null hypothesis) at the 1% level and * indicates significance at the 5% level. Below the calculated values of the estimated coefficients by the Johansen cointegration method are the standard errors of the estimates. From the k-order VAR model, the ∆Xt and Xt-k are regressed on a constant and ∆Xt-1, …, ∆Xt-k+1. The obtained residuals R0t and Rkt are used in the construction of the residual product matrices matrix Sij. The matrix of cointegrating vectors is then estimated as the eigenvectors associated with the eigenvalues λ1 > λ2 > … > λ r > 0 found as the solution to ⏐λSkk - Sk0S00–1S0k⏐. The test statistics (for n=4 variables in the system) are based on the maximal eigenvalue test and the trace test. The maximal eigenvalue test (null is r cointegration vectors (C.V.) against the alternative of r+1 cointegration vectors) is based on the statistic: λmax = - Tln(1λ r+1), where T is the sample size, r is the number of cointegrating vectors, and λ i are the eingenvalues above. The trace test (null is at most r C.V.’s against the alternative of more than r C.V.’s) is based on the statistic: P λ trace = -T Σ ln(1- λ i). The lag-length selection criteria used were given combination of sequential modified likelihood i=r+1 ratio (LR) test, the final prediction error (FPE), and the Akaike (AIC), Schwarz (SIC) and Hannan-Quinn (HQIC) information criteria and Lagrange Multiplier (LM) serial correlation tests summarized at the rightmost column. Below the LM serial correlation statistics in brackets is the p-value associated with the null of no serial correlation at lags 4 or 8.

32 Table 5. Cointegration Tests in Brownsville of Full Model in Logs: [L, wages, VAS, LNAT] Null Hyp. VAR Lag Trace 5% 5% Max. Model Specification and Critical on Coint. Length / Eigenv. Critical Test Estimation of Johansen Value Vectors LM Test Value Test Cointegration Vectors FULL MODEL – CITY 46.30** 27.07 70.16** 47.21 3 by SIC No C.V.s Log(L)= -4.129 log(WTP) 15.84 20.97 23.87 29.68 At most 1 (0.436) 7.95 14.07 8.03 15.41 At most 2 12.91 16.50 0.08 3.76 0.08 3.76 At most 3 [0.68] [0.42] + 0.141 log(VAS) (0.024) +3.016 log(LNAT) (0.209) FULL MODEL - MANUFACTURING

Log(LM)= -2.449 log(WTP) (0.888) -0.059 log(VAS) (0.044) +0.602 log(LNAT) (0.420)

42.28** 21.30* 11.08 4.27

27.07 20.97 14.07 3.76

78.93** 36.65** 15.35 4.27*

47.21 29.68 15.41 3.76

No C.V.s At most 1 At most 2 At most 3

4 by SIC

34.83** 22.63* 2.20 1.40

27.07 20.97 14.07 3.76

61.07** 26.23 3.61 1.40

47.21 29.68 15.41 3.76

No C.V.s At most 1 At most 2 At most 3

6 by FPE, AIC, HQ

33.63** 17.37 4.91 1.34

27.07 20.97 14.07 3.76

57.25** 23.62 6.25 1.34

47.21 29.68 15.41 3.76

No C.V.s At most 1 At most 2 At most 3

3 by SIC

23.29 13.81 [0.11] [0.61]

FULL MODEL – TRADE

Log(LT)= -13.320 log(WMP) (2.154) + 0.159 log(VAS) (0.083) + 3.913 log(LNAT) (0.641)

12.51 10.66 [0.71] [0.83]

FULL MODEL - GOVERNMENT

Log(LG)= +0.651 log(WTP) (0.761) -0.003 log(VAS) (0.040) +1.550 log(LNAT) (0.360)

20.70 25.25 [0.19] [0.07]

Notes: Intercept and trend are included in cointegrating equation and test VAR. The data in levels have linear trends but the cointegrating equation has only intercepts. The symbols ** indicates significance (rejection of the null hypothesis) at the 1% level and * indicates significance at the 5% level. Below the calculated values of the estimated coefficients by the Johansen cointegration method are the standard errors of the estimates. From the k-order VAR model, the ∆Xt and Xt-k are regressed on a constant and ∆Xt-1, …, ∆Xt-k+1. The obtained residuals R0t and Rkt are used in the construction of the residual product matrices matrix Sij. The matrix of cointegrating vectors is then estimated as the eigenvectors associated with the eigenvalues λ1 > λ2 > … > λ r > 0 found as the solution to ⏐λSkk - Sk0S00–1S0k⏐. The test statistics (for n=4 variables in the system) are based on the maximal eigenvalue test and the trace test. The maximal eigenvalue test (null is r cointegration vectors (C.V.) against the alternative of r+1 cointegration vectors) is based on the statistic: λmax = - Tln(1λ r+1), where T is the sample size, r is the number of cointegrating vectors, and λ i are the eingenvalues above. The trace test (null is at most r C.V.’s against the alternative of more than r C.V.’s) is based on the statistic: P λ trace = -T Σ ln(1- λ i). The lag-length selection criteria used were given combination of sequential modified likelihood i=r+1 ratio (LR) test, the final prediction error (FPE), and the Akaike (AIC), Schwarz (SIC) and Hannan-Quinn (HQIC) information criteria and Lagrange Multiplier (LM) serial correlation tests summarized at the rightmost column. Below the LM serial correlation statistics in brackets is the p-value associated with the null of no serial correlation at lags 4 or 8.