1

What Explains Unemployment in U.S. - Mexican Border Cities?

André Varella Mollick∗ TP

PT

Department of Economics and Finance College of Business Administration University of Texas - Pan American 1201 W. University Dr. Edinburg, TX 78539-2999, USA E-mail: [email protected] Tel.: +1-956-316-7913 and fax: +1-956-384-5020. HTU

UTH

Abstract: The unemployment rate of U.S.-Mexico border cities has stood remarkably higher than the U.S. average. Using annual data from 1990 to 2005, we contrast large border MSAs (Brownsville, McAllen, Laredo, and El Paso in Texas and El Centro in California) to a panel of MSAs in the same states (Austin, Dallas, Houston, San Antonio, Los Angeles and San Francisco). Focusing on the industry composition of employment and population growth, we report several panel data results confirmed by error correction adjustment. First, the national unemployment rate does not help explain the local border cities unemployment but does so for the panel of large MSAs. Second, the relative employment indices have statistically significant effects only for the border panel: increases in employment concentration within an industry lead to higher local border unemployment. Third, higher population density lowers unemployment for border cities.

Keywords: Border, Employment Decomposition, Industrial Diversity, MSA, Unemployment Rate. JEL Classification Numbers: J6, R1, R2.

∗

TP

Two anonymous referees of this journal provided useful comments on a first draft. The usual disclaimer applies. PT

2

1. Introduction Following the seminal papers by Thirlwall (1966) and Brechling (1967) for the U.K. and by Blanchard and Katz (1992) on U.S. states, several authors have estimated the effects of the national unemployment rate on local unemployment rates. Examples include: Decressin and Fatás (1995) for Europe versus the U.S., Martin (1997) and Gray (2004) for British regions, Jimeno and Bentolila (1998) for Spain, and Galiani et al. (2005) for Argentina. Several reasons justify the regional approach to labor market conditions. As stated by Decressin and Fatás (1995), the national labor market is a fairly arbitrary aggregation of several heterogeneous dynamics. Further, the region-specific shocks are known to trigger different adjustment mechanisms than national shocks. More generally, common currency considerations inspired in Mundell (1961) suggest we explore whether labor market disturbances are distributed more or less symmetrically across the regions. This paper studies unemployment along the U.S.-Mexican border. There are three major methodological motivations associated with our research. First, we focus on the five largest metropolitan statistical areas (MSAs) along the U.S.-Mexican border and compare their trends with the largest MSAs in the same state located away from the border. Using annual data from 1990 to 2005, we examine two panels of large MSAs: one composed of border cities (Brownsville, McAllen, Laredo, and El Paso in Texas and El Centro in the “Imperial Valley” of California) and another of large metropolitan areas in the same states (Austin, Dallas, Houston, San Antonio, Los Angeles and San Francisco). While the unemployment figures are consistently higher than the nation for border cities, large cities’ unemployment follow the national average closely.1 1

These contrasting patterns may suggest some sort of “thick labor markets hypothesis” holds: when the cities are sufficiently large and developed unemployment is low, and vice versa. See Neumann and Topel (1991) and Gan and Zhang (2006) for formal approaches.

3

Second, there is the model specification issue. The national unemployment rate is far from being the major explanatory factor for local unemployment, especially for border cities. Since theories on the relatively high unemployment rate of border cities have not yet been found to be particularly satisfactory - e.g., seasonality suggested for California’s Imperial County by EDD of California (2005) - we focus herein on two sets of potential explanations for local unemployment: the industry composition of employment and population growth.2 The first of these follows Simon (1988), Izraeli and Murphy (2003) and Mizuno et al. (2006) who define Herfindahl indexes for a metropolitan area and expect a positive relationship between the degree of concentration and local unemployment rates. Following Duranton and Puga (2000), we complement the Herfindahl index with specialization indexes, adjusted for national trends. The second systematic source of unemployment changes is related to the behavior of population density in local areas, since city size has been shown to systematically affect local unemployment rates as argued by Vipond (1974), Sirmans (1977) and Alperovich (1993). We expect population density (POPDEN) reduces the unemployment level: higher density means lower production costs (through lower transportation and communication costs), which facilitates the matching between firms and workers. Third, the empirical methodology in this paper fully explores the statistical content of the long-run relationships. We follow Marston (1985), who viewed local unemployment as two basic situations: the disequilibrium one due to high costs of migration and the slow labor flow between areas and the equilibrium one such that workers migrate until there is no further incentive to move. We implement the empirical 2

What came to be called the “mystery of regional unemployment” has been surveyed by Elhorst (2003), who reviews 41 empirical studies and identifies 13 sets of explanatory variables. His conclusion is that industry diversity indices are more suitable to explain local unemployment: “regions with diverse sources of employment are likely to provide greater opportunities for labour redeployment between firms and industries in response to their changing employment needs.” Elhorst (2003, p. 735). See also Partridge and Rickman (1997) for U.S. state unemployment with a very comprehensive list of independent variables.

4

methodology based on the Engle and Granger (1987) error correction model (ECM), in which “adjustment occurs over the long run but in the short run barriers to migration allow the disequilibrium to exist for some time.” Marston (1985, p. 60). Our major results are as follows. First, the national unemployment rate turns out to be not important for explaining the local unemployment of border cities but very much so for the panel of large MSAs. Second, the relative employment indices have statistically significant effects only for the border panel of cities. As employment concentration in an industry increases, local unemployment rates increase: the relative diversity index coefficient varies from 1.88 to 2.09. Third, higher population density stimulates economic activity and lowers the unemployment rate for border cities (coefficient ranges from -0.84 to -1.77), consistent with Alperovich (1993) for Israel and B

B

with Izraeli and Murphy (2003) for U.S. states. Fourth, the ECM shows that a large part of the lagged equilibrium deviations are adjusted in the next period.

2. The Data The unemployment rate and employment figures at the MSA and national level come from the U.S. Bureau of Labor Statistics (BLS) website (http://www.bls.gov). The HTU

UTH

unemployment rate is based on the monthly sample survey Current Population Survey on about 60,000 households. The BLS website contains details on the sampling methodology. While these figures are the most widely used unemployment statistics for the U.S., problems include the treatment of discouraged workers, part-time workers, unreported HT

TH

HT

TH

legal employment, and unreported illegal employment. On the basis of population and geographic location, we select the metro urban areas as follows (codes for graphical analysis appears in parenthesis): Austin-Round Rock (au), Brownsville-Harlingen (b), Dallas-Fort Worth-Arlington (dal), El Paso (el), Houston-Sugar

5

Land-Baytown (hou), Laredo (lar), McAllen-Edinburg-Mission (mc) and San Antonio (san) in Texas; and El Centro (centro), Los Angeles-Long Beach-Santa Ana (la), San Diego-Carlsbad-San Marcos (sd), and San Francisco-Oakland-Fremont (sf) in California. The choice of cities follows from contrasting the two sets of MSAs and also because there are no MSAs in Arizona and New Mexico bordering Mexico. The time period (1990 to 2005) is chosen for data availability for the employment breakdown. We define three different panels: one for the 5 border cities (5-border), one for the very large MSAs located away from the border (6-large), and one for all cities, including San Diego (all 12 cities). Since San Diego is a border city and is also a very large metropolitan area, we omit it from the panel based approach. Figure 1 shows local unemployment rates at the five border MSAs compared to the national unemployment rate (Un). It is clear that all border cities have much higher unemployment rates than the nation, with some cities (Imperial Valley in California and McAllen-Edinburg-Mission in Texas) recording higher than 20% rates in the early to mid 1990s. It is visible the decrease in the unemployment rates of the border cities, following the economic boom of the late 1990s. After the recession of 2001-2002, local unemployment rates edge higher along with Un. Yet, the nation’s unemployment rate seems to work as a lower bound on the local unemployment rates of these cities for the overall time period.3 [Figures 1 and 2 here] Some convergence to Un is also observed at large MSAs in California and Texas in Figure 2. However, contrary to the border cities, some cities (Austin, Dallas and San Francisco) have lower unemployment rates than Un over the sample. Several measures in the statistical analysis are calculated from the employment breakdown of total nonfarm employment of all employees at the MSA level. Ten industry 3

Some attribute the boom at the U.S. - Mexico border as a major factor propelling the Texas economy to outperform the U.S. economy in the 1990s. See, for example, Cañas et al. (2005).

6

sectors are examined as follows (codes in parenthesis): natural resources, mining & construction (min), manufacturing (man), trade, transportation & utitlities (tra), information (inf), financial activities (fin), professional and business services (pro), education and health services (edu), leisure and hospitality (lei), other services (serv), and government (gov).4 Duranton and Puga (2000) propose to focus on the employment share of each city’s largest sector. If we denote sij as the share of industry j in city i, we can define the specialization index:

ZIi = Max (sij) j

(1).

Since certain sectors account for a larger share of national employment than others, Duranton and Puga (2000) propose a correction based on a city’s relative (rather than absolute) specialization:

RZIi = Max (sij / sj) j

(2),

where sj is the share of industry j in national employment. We will refer to (2) as the relative specialization index (RZI). In our sample of 12 cities, the sectors with highest share of employment were government and trade transportation and utilities. The highest of all these RZI’s was calculated for the government sector in the El Centro MSA: increasing from 1.889 in 1990 to 2.425 in 2005. There are also downward trends in RZI for 4

The BLS dataset provides further breakdown of manufacturing employment for large MSAs but does not do so for cities located along the U.S.-Mexico border. For this reason, we are unable to provide a finer calculation of indices based on the several components of durable sectors, for example, as done by Henderson (1997a) for 20 manufacturing sectors.

7

Austin and San Antonio. If the specialization thesis is correct, cities that have its index increase over time should have the local unemployment rate decrease, ceteris paribus. Following Simon (1988), Izraeli and Murphy (2003) and Mizuno et al. (2006), for N industries the Herfindahl index of a metropolitan area i is defined as follows:

N

Σ s2ij

Hi ≡ B

B

P

PB

(3)

B

j=1 The Herfindahl index calculated by (3) for the 10 industries listed show different patterns across panels. According to the diversity thesis, as the degree of concentration increases (H approaches 1), the local unemployment rate should increase since the structural unemployment component tends to be smaller, ceteris paribus. Conversely, when H takes the minimum of 1/N the local unemployment rate should be smaller. As recommended by Duranton and Puga (2000), we adjust H by computing a relative diversity index (RDI). We sum for each city, over all sectors, the absolute value of the difference between each sector’s share in local employment and its share in national employment:

RDIi = Σ │sij - sj│ j B

B

B

B

B

B

(4)

An increase in RDI means more employment concentration; a decrease in RDI means more diversity in employment. In our sample, RDI changed upwards from 0.622 to 0.689 in El Centro over the period, as well as in Brownsville, Laredo and McAllen; and downwards in El Paso. In contrast, all non-border cities display a decrease in the index

8

over the 16-year period, with Austin yielding the most notable change towards diversification of industry employment from 0.253 in 1990 to 0.149 in 2005. Table 1 contains descriptive statistics for large and border panels of MSA population growth during the 1990s. We calculate the series POPDEN as the population density: total county population divided by the square miles of the largest county. Other than the county of Austin which grows by 41% (48% by the whole MSA), population growth in border MSAs is usually much higher as both Laredo and McAllen MSAs have close to 50% population growth in the decade. El Centro and Brownsville urban areas come next with close to 30% growth decades in a single decade. We conjecture the composition of industry employment and population density help explain local unemployment rates. An omitted factor, due to data availability at the MSA or county level, is net migration as captured by Evans and McCormick (1994). We argue that, in growth form, population density should be capturing net migration, a critical feature for U.S.-Mexican border urban areas. [Table 1 here]

3. The Empirical Methodology Given that the unemployment rate is a bounded series in the [0, 1] interval, over the very long term unemployment rates should not follow I (1) random-walk processes with drift. Under relative short time spans, however, it may be the case that U follows an untrended I (1) process, having a non-stationary degree of persistence, as discussed by León-Ledesma and McAdam (2004). Empirical evidence on local unemployment has been mixed and can be verified in Chapman (1991), Jimeno and Bentolila (1998), Payne et al. (1999), Maki-Arvela (2003), Gray (2004), and Galiani et al. (2005).

9

We consider panel unit root tests and incorporate lagged first-differences to account for serial correlation in the unemployment rate series. The equation below is estimated for each of the panels discussed earlier:

k ∆uit = α0 + α1uit-1 + Σ αij ∆ui t-j + νit j=1 B

B

B

B

B

B

B

B

B

B

B

B

B

(5),

B

where: uit is the unemployment rate in MSA i at time t, ∆ is the first-difference operator, B

B

and k is the number of lags. We report below panel unit roots proposed by Levin, Lin and Chu (2002), denoted LLC, and Im, Pesaran and Shin (2003), denoted IPS, with the Schwarz criterion employed for lag-length selection. The null hypothesis of unit root is α1 = 0; failure to reject the null is evidence in support of a unit root in the series. More B

B

powerful than these tests is the recent panel-LM unit-root test by In et al. (2005), which is robust to structural shifts in the data. In our context, however, a 16-year time span is not that long to justify strong structural breaks in the series. The empirical methodology estimates the following panel data model for local unemployment rates (Uit): B

B

Uit = β0i + β1 trend + β2 Unt + νit B

B

B

B

B

B

B

B

B

B

B

(6),

B

where: the parameter β0i represents unobserved city specific fixed effects; trend captures B

B

the time trend and Unt is the national unemployment rate. The coefficient β2 is expected B

B

B

B

to be positive in (6). The fixed effects control for factors that vary across cities but are time invariant. It is likely that there are unobserved city specific effects that have a systematic impact on local unemployment rates. These include amenities, cultural background of the population, persistent tastes for a specific sort of job within a locality,

10

and so on. Previous applications of the fixed effects model to regional unemployment include Izraeli and Murphy (2003) and Bockerman (2003). The individual fixed effects may be either assumed to be correlated with the right hand side variables (fixed effects model: FEM) or be incorporated into the error term (random effects model: REM) and assumed uncorrelated with the explanatory variables.5 In order to ease interpretation of the coefficients, we take logarithms on the unemployment rates in both sides of the equation. We focus herein on two specific sets of potential explanations for unemployment fluctuations. The first follows Simon (1988), Izraeli and Murphy (2003) and Mizuno et al. (2006) who define Herfindahl indexes for a metropolitan area for N industries. A positive relationship is expected between the degree of concentration (H) and local unemployment. The second systematic source of unemployment changes is population density. The size of the metropolitan area has been shown to systematically affect local unemployment rates as argued by Vipond (1974), Sirmans (1977) and Alperovich (1993). In general, the larger the city size the more employment opportunities. However, Vipond (1974) stresses that larger cities have wider wage dispersion and job seekers may wish to prolong job search. Alperovich (1993) finds, for Israeli urban areas over 1972-1983, the sign of city size to be strongly negative. We follow Izraeli and Murphy (2003) and consider population density (POPDEN), which is expected to have a negative effect on the unemployment level. Higher density means lower production costs (lower transportation and communication costs), which helps the employment matching. These two forces lead us to the augmented model below:

5

We perform Hausman tests on the product of the difference between the parameter vector estimated by FEM and the vector estimated by REM and the covariance of the difference. As discussed in Greene (2003), the test is no longer valid when the matrix difference is not positive definite. We conduct both set of estimations and report below the results from the FEM, with REM results discussed in earlier versions.

11

Uit = β0i + β1 trend + β2 Unt + β3 INDEXit + β4 POPDENit + νit B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

(7),

where: INDEX (in logarithms) may be either RZI for the relative specialization index or RDI for the relative diversity index (Herfindahl); and POPDEN is the log of the population density. Exploring the panel data structure, we estimate (6) and (7) using the feasible generalized least squares (FGLS) fixed-effects model with seemingly unrelated regression (SUR) weights, when the residuals are both cross-section heteroskedasticity and contemporaneously correlated.6 Differencing (7) successfully remove serial correlation TP

PT

problems.7 In addition to (6) and (7), we consider an ECM that combines the short-run TP

PT

properties of statistical relationships in first-differenced form with the long-run properties in level form, as in Engle and Granger (1987):

∆(U)it = α + β0i + β1 ECMit-1 + β2∆(Un)t-1 + β3 ∆(INDEX)it + B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

+ β4 ∆(POPDEN)it + β5∆(U)it-1 + νit B

B

B

B

B

B

B

B

B

B

(8),

where ECMit-1 are lagged one-period residuals from each state level equation (6) or (7) B

B

with different (RHS) specifications. We expect the β1-coefficient in (8) to be negative in an B

B

error-correction fashion. Estimating (8) by FGLS methods, we then use the SUR method for calculating the weights of the variance-covariance matrix, which is appropriate when all regressors are exogenous and the errors are heteroskedastic and contemporaneous correlated. The weights and coefficients are updated continuously until convergence. 6

The alternative is to use the generalized method of moments (GMM) estimator as in Henderson (1997b) for U.S. manufacturing and in Galiani et al. (2005) for regional unemployment rates of Argentina. Application of the FGLS on a lagged dependent variable may result in biased estimates if the contemporaneous error term is correlated with any time average of the lagged dependent variable. See Henderson (1997b) for details. In the present setting, there is no lagged dependent variable and we make use of the result in Pirotte (1999), who shows that the probability limit of the between estimator of a static relation converges to the long run effects. TP

PT

12

4. Empirical Results Estimates of (5) for LLC and IPS panel unit roots for the three different panels are performed. For unemployment rates, the unit root null is rejected at 1 or 5% significance levels for all panels, except for the border cities panel under the IPS test. When allowing for the time trend, the results are mixed for the border cities panel: only LLC test rejects the null. This casts doubt on the stationarity of the unemployment rates. For regional RZI and RDI measures, the null is rejected for LLC tests but not for IPS tests. For population density, the panel unit roots always reject the null. Table 2 reports the estimates of the simple model with only the time trend and the national unemployment levels as regressors. The method of estimation is FEM with SUR cross-section weights. As conjectured above, local unemployment rates respond positively to national unemployment rates: when the trend is included β2 = 0.766 for the panel of all B

B

cities and β2 =1.189 for the panel of large and non-border cities. The latter response, in B

B

particular, suggests a more than proportional effect for the large MSAs: when the national unemployment rate grows by 1%, the local unemployment rate responds by 1.2%. When the time trend is included, the exception is the border cities panel when the national unemployment rate is not statistically significant in predicting local unemployment fluctuations. The estimates of (6) corroborate Figures 1 and 2: local unemployment follows national unemployment more closely for the non-border cities. The formal tests for serial correlation suggest misspecifications all over Table 2. The close to 1 adjusted R2 may also imply spurious regressions, which can be addressed in P

P

the panel cointegration framework. If the residuals to (6) are stationary, the ECM equation must indicate the extent of the deviations away from the long-run equilibrium. The results 7

We conduct two serial correlation tests derived from the Lagrange Multiplier (LM) Breusch-Godfrey test. For each panel equation, the computed residuals are regressed on the model’s independent variables and on the (lagged one period) residuals. TP

PT

13

in Table 2 always provide stationary residuals for the panel of large MSAs and also for the border cities. [Table 2 here] We turn next to the augmented models for the unemployment rate that take into account industry-mix employment composition as well as population changes. The estimation of equations separately for RZI and RDI, as well as with both explanatory variables estimated jointly, allows inference to be made on interaction effects. Table 3 reports several interesting findings for three specifications of equation (7). First, the effect of the national unemployment rate is strong and more than proportional for the large cities, varying from β2 = 1.153 to β2 = 1.176 across columns (4) – (6). For the border cities, the B

B

B

B

coefficient on the national unemployment rate is never statistically significant. Second, the relative indices have statistically significant effects only for the border panel, varying from β4 = 1.879 to β4 = 2.089 for the relative diversity index. As concentration increases, local B

B

B

B

unemployment rates increase as conjectured theoretically. There is a weaker effect (significant at 5%) for the relative specialization index: as specialization increases the local unemployment rate falls (β3 = -0.513), also in agreement with the conjecture. B

B

Third, the effect of population density on local unemployment is remarkably different across panels: for the border cities, higher population density stimulates economic activity and lowers the unemployment rate (β5 ranges from -0.837 to -1.765). This is B

B

consistent with the results in Alperovich (1993) for Israel and with Izraeli and Murphy (2003) for U.S. states. On the other hand, the sample of large cities presents the reverse result: higher population density reduces economic activity and increases the unemployment rate (β5 ranges from 0.799 to 0.823). The latter suggests congestion effects B

B

and diseconomies associated with large metro areas. [Table 3 here]

14

Since the residuals underlying the several specifications of (7) are stationary according to both panel unit roots as shown at the bottom of Table 3, an ECM combines the short-run properties of statistical relationships in first differences with the long-run properties of the model in level form. We estimate (8) using FGLS methods as done for models (6) and (7). The lagged residuals from each equation in levels, denoted ECMit-1, are B

B

introduced as an explanatory variable in an error-correction fashion. The results of these estimations are reported in Table 4. The coefficient for ECMit-1, B

β3, is negative and significant in all specifications: as deviations from the long-run B

B

B

equilibrium increase, the impact on the dependent variable decreases in an error-correction fashion. These estimates show that between 48% (column 2) and 61% (column 3) of the lagged long-run equilibrium deviations are adjusted in the next period for the border cities panel and about 43-45% of the lagged long-run equilibrium deviations are adjusted in the next period for the non-border cities panel. The negative and strongly significant figures imply a fast adjustment to the long-run equilibrium and reinforce the previous long-run results.8 TP

PT

[Table 4 here] The estimated coefficients on the β2-coefficient associated with the national B

B

unemployment rate are again strong and statistically significant for the panels with the large MSAs. The coefficients are close to 1 for the non-border cities panel. They are, however, much weaker, ranging between 0.308 and 0.354, for the border cities panel. There is thus for the ECM model a similar relationship than the one detected earlier between local and national unemployment rates. Variations in the relative indices have no

TP

8

The only article we know of reporting an error correction term is Martin (1997) with β1-coefficients for the U.K. over 1965-1995 varying between -0.35 and -0.56. The underlying model in his case, however, is (6) with only national unemployment rates. Besides, Martin (1997) looked at regional unemployment from the viewpoint of time series and not from a panel data structure. PT

B

B

15

significant explanatory role in local unemployment fluctuations away from the long-run equilibrium. Away from the steady-state, increases in population density lead to higher local unemployment for border cities. The magnitude of the estimated coefficient is very large (between 8.44 and 9.97), suggesting an amplified response to population density. Increases in population density lead to higher local unemployment for border cities. One possible interpretation is that positive net migration rate pushes the unemployment rate higher when the economy is away from the long-run equilibrium. Therefore, areas with positive net migration rates (captured by higher population growth due to lack of data at the MSA or county level) have their unemployment rate move higher when the economy is away from the long-run equilibrium. Finally, the explanatory power of the ECM models is very good, especially for the non-border city panels. The adjusted R2 statistics range from about 40% (5 border cities) to about 80% (6 large cities). Neither the DW statistics nor the more formal LM tests detect serial correlation, which suggests a very good fit for the ECM model.

5. Concluding Remarks Rather than repeating the main findings, we address avenues for future research. In addition to net migration discussed in Evans and McCormick (1994), at least three extensions of this work are worth exploring. First, Hanson (2001) has reported that the activity of Mexican maquiladora production has a large impact on U.S. cities close to the border. Mollick et al. (2006) have revisited this issue for Brownsville-Matamoros and El Paso-Juárez city pairs and suggest the impact is not confined to manufacturing. Incorporating this finding into local unemployment behavior is part of our research agenda. Second, city-specific effects may, in fact, influence labor types differently, as

16

well as our industry indices may have different effects on skilled or unskilled unemployment. Examination of this issue may provide further insights into the behavior of local unemployment. Third, border cities may serve as temporary destinations for immigrants, which may make the natural unemployment rate in border cities be higher than in non-border cities. Controlling for the number of unemployed migrants, however, is only possible through survey data.

17

References Alperovich G (1993) City size and the rate and duration of unemployment: Evidence from Israeli data. Journal of Urban Economics 34: 347-357. Blanchard O, Katz L (1992) Regional evolutions. Brooking Papers of Economic Activity 1: 1-75. Bockerman P (2003) Unravelling the mystery of regional unemployment in Finland. Regional Studies 37 (4): 331-340. Brechling F (1967) Trends and cycles in British regional unemployment. Oxford Economic Papers 19: 1-21. Cañas J, Coronado R, and Gilmer R (2005) Texas Border, Employment and Maquiladora Growth, The Face Texas: Jobs, People, Business and Change, October, Federal Reserve Bank of Dallas. doi: http://www.dallasfed.org/research/pubs/fotexas/fotexas_canas.pdf Chapman P (1991) The dynamics of regional unemployment in the U.K., 1974-1989. Applied Economics 23: 1059-1064. Decressin J, Fatás A (1995) Regional labor market dynamics in Europe. European Economic Review 39: 1627-1655. Duranton G, Puga D (2000) Diversity and specialization in cities: Why, where and when does it matter? Urban Studies 37 (3): 533-555. Elhorst JP (2003) The mystery of regional unemployment differentials: Theoretical and empirical explanations. Journal of Economic Surveys 17 (5): 709-748. Employment Development Department (EDD) of the State of California, 2005. Imperial County: Snapshot. doi: http://www.labormarketinfo.edd.ca.gov/ Engle R, Granger C (1987) Cointegration and error correction: Representation, estimation, and testing. Econometrica 55 (2): 251-276. Evans P, McCormick B (1994) The new pattern of regional unemployment: Causes and policy significance. Economic Journal 104: 633-647. Galiani S, Lamarche C, Porto A, and Sosa-Escudero W (2005) Persistence and regional disparities in unemployment (Argentina 1980-1997). Regional Science & Urban Economics 35: 375-394.

18

Gan L, Zhang Q (2006) The thick market effect on local unemployment rate fluctuations. Journal of Econometrics 133 (1): 127-152. Gray D (2004) Persistent regional unemployment differentials revisited. Regional Studies 38 (2): 167-176. Greene WH (2003) Econometric analysis. 5th Edition. Prentice-Hall: Upper Saddle River, NJ. Hanson G (2001) U.S. – Mexico integration and regional economies: Evidence from border-city pairs. Journal of Urban Economics 50: 259-287. Henderson V (1997a) Medium size cities. Regional Science & Urban Economics 27: 583-612. Henderson V (1997b) Externalities and industrial development. Journal of Urban Economics 42: 449-470. Im KS, Lee J, Tieslau M (2005) Panel LM unit-root tests with level shifts. Oxford Bulletin of Economics and Statistics 67 (3): 393-419. Im KS, Pesaram MH, Shin Y (2003) Testing for unit roots in heterogenous panels. Journal of Econometrics 115 (1): 53-74. Izraeli O, Murphy, KJ (2003) The effect of industrial diversity on state unemployment rate and per capita income. Annals of Regional Science 37: 1-14. Jimeno J, Bentolila S (1998) Regional unemployment persistence (Spain, 1976-1994). Labor Economics 5: 25-51. León-Ledesma M, McAdam P (2004) Unemployment, hysteresis and transition. Scottish Journal of Political Economy 51 (3): 377-401. Levin A, Lin CF, Chu CSJ (2002) Unit roots in panel data: Asymptotic and finite sample properties. Journal of Econometrics 108: 1-24. Maki-Arvela P (2003) Regional evolutions in Finland: Panel data results of a VAR approach to labour market dynamics. Regional Studies 37 (5): 423-443. Marston S (1985) Two views of the geographic distribution of unemployment. Quarterly Journal of Economics: 57-79. Martin R (1997) Regional unemployment disparities and their dynamics. Regional Studies 31 (3): 237-252.

19

Mizuno K, Mizutani F, and Nakayama N (2006) Industrial diversity and metropolitan unemployment rate. Annals of Regional Science 40 (1): 157-172. Mollick AV, Cortez A, and Olivas R (2006) Local labor markets in U.S.-Mexican border cities and the impact of Mexican maquiladora production. Annals of Regional Science 40: 95-116. Mundell R (1961) A theory of optimum currency areas. American Economic Review 51: 657-664. Neumann G, Topel, G (1991) Employment risk, diversification, and unemployment. Quarterly Journal of Economics 106: 1341-1365. Partridge M, Rickman D (1997) The dispersion of U.S. state unemployment rates: The role of market and non-market equilibrium factors. Regional Studies 31 (6): 593-606. Payne J, Ewing B, and George, E (1999) Time series dynamics of U.S. state unemployment rates. Applied Economics 31: 1503-1510. Pirotte A (1999) Convergence of the static estimation toward the long-run effects of dynamic panel data models. Economics Letters 63: 151-158. Simon CJ (1988) Frictional unemployment and the role of industrial diversity. Quarterly Journal of Economics 103 (4): 715-728. Sirmans CF (1977) City size and unemployment: Some new estimates. Urban Studies 14: 99-101. Thirlwall A (1966) Regional unemployment as a cyclical phenomenon. Scottish Journal of Political Economy 13: 205-219. Vipond J (1974) City size and unemployment. Urban Studies 11 (1): 39-46.

20

Figure 1. Unemployment Rates at Border MSAs and in the U.S. (Un). Unemployment Rates at Border MSAs versus the Nation (Un) 35.0

30.0

25.0

20.0

15.0

10.0

5.0

0.0 1990

1991

1992

1993

1994 Un

1995

1996

U_centro

1997 U_b

1998 U_el

1999

2000 U_lar

2001 U_mc

2002

2003

2004

2005

21

Figure 2. Unemployment Rates at Non-Border MSAs and in the U.S. (Un).

Unemployment Rates at Large Non-Border MSAs versus the Nation (Un) 10.0

9.0

8.0

7.0

6.0

5.0

4.0

3.0

2.0

1.0

0.0 1990

1991

1992

1993

1994 Un

1995 U_la

1996 U_sf

1997 U_au

1998

1999 U_dal

2000 U_hou

2001

2002

U_san

2003

2004

2005

22

Table 1: Population Growth Rates in MSAs and in their Largest Counties. 1990 MSA pop.

2000 MSA pop.

Decade MSA pop. growth rate

1990 largest county pop.

2000 largest county pop.

Decade largest county pop. growth rate

Largest County square miles

Austin-Round Rock

846,000

1,250,000

48%

576,407

812,280

41%

989

Dallas-Fort WorthArlington

3,989,000

5,162,000

29%

1,852,810

2,218,899

20%

902

Houston-Sugar Land-Baytown

3,767,000

4,715,000

25%

2,818,199

3,400,578

21%

1778

San Antonio

1,408,000

1,712,000

22%

1,185,394

1,392,931

18%

1248

Los AngelesLong BeachSanta Ana

11,274,000

12,366,600

10%

8,863,164

9,519,338

7%

4079

San FranciscoOaklandFremont

3,684,000

4,124,000

12%

1,185,394

1,392,931

7%

91

San DiegoCarlsbad-San Marcos

2,498,000

2,814,000

13%

2,498,000

2,814,000

13%

4281

Border MSAs BrownsvilleHarlingen

260,000

335,000

29%

260,000

335,000

29%

905

El Paso

592,000

680,000

15%

592,000

680,000

15%

1057

Laredo

133,239

193,117

45%

133,239

193,117

45%

3363

McAllenEdinburgMission

384,000

569,000

48%

384,000

569,000

48%

1596

El Centro

109,303

142,361

30%

109,303

142,361

30%

1057

Large MSAs

Sources: U.S. Census Bureau and author’s calculations.

23

Table 2: FGLS Estimations of Unemployment in Levels. Uit = β0 + β2 Un + εit Uit = β0 + β1 trend + β2 Un + νit B

B

B

0.421*** (0.019)

B

β1

B

B

B

B

B

B

B

B

B

B

B

B

0.796*** (0.027)

1.656*** (0.297)

-0.013*** (0.001)

B

β2

B

B

B

B

B

B

B

(6)

B

5-Border Cities Panel

12-Cities Panel β0

B

6-Large Cities Panel 2.885*** (0.425)

-0.104 (0.088)

-0.047*** (0.005)

-0.509*** (0.128)

0.014*** (0.002)

0.929*** (0.012)

0.766*** (0.015)

0.518*** (0.166)

0.003 (0.225)

1.011*** (0.050)

1.189*** (0.069)

1.910

1.850

1.164

1.168

1.257

1.167

Adj. R2

0.9999

0.9997

0.989

0.986

0.988

0.987

LM t-stat

30.668***

31.364***

B

B

DW P

P

LM NR2 stat

16.694***

11.070***

14.450***

16.736***

144.90***

148.68***

53.55***

40.65***

52.74***

52.02***

LLC

-0.524 [0.300]

-0.737 [0.231]

1.093 [0.863]

-2.447*** [0.007]

-1.293* [0.098]

-2.323** [0.010]

IPS

-0.916 [0.180]

-1.071 [0.142]

2.010 [0.978]

-1.642** [0.050]

-1.691** [0.045]

-2.986*** [0.001]

P

P

Panel Unit Root Tests on Residuals

Notes: Logarithms are taken on all unemployment rate series. The total number of observations is either 80 for the border sample or 96 for the sample of large cities (5 or 6 local areas times 16 years). For the 12 cities panel, it is 192 (12 local areas times 16 years). The entries below the coefficients are White-cross section standard errors corrected for degrees of freedom. The LM statistic (null of no serial correlation) is the value associated with the lagged residual computed in the auxiliary regression. It follows a chi-squared distribution with degrees of freedom equal to the number of parameters (q): χ2 (2) = 5.99; χ2 (3) = 7.81; and χ2 (4) = 9.49. The symbols *, **, and *** refer to levels of significance of 10%, 5%, and 1%, respectively. For the panel unit root tests LLC and IPS, the p-values are given in brackets.

24

Table 3: FGLS Logarithmic Estimations of Unemployment: Relative Indices. Uit = β0 + β1 trend + β2 Unt + β3 INDEXit + β4 POPDENit + νit B

B

B

B

B

B

B

B

B

B

B

B

B

B

5-Border Cities Panel β1

(trend)

B

B

(Un)t

β2 B

B

B

B

B

B

(RZI)t

β3 B

B

B

-0.030*** (0.005)

-0.014*** (0.004)

0.069 (0.180)

0.123 (0.182)

(RDI)t

β4

B

B

B

B

B

(7)

6-Large Cities Panel

0.367 (0.230)

B

B

-0.0003 (0.008)

-0.001 (0.002)

0.0001 (0.002)

-0.001 (0.004)

0.028 (0.197)

1.176*** (0.074)

1.156*** (0.074)

1.153*** (0.077)

-0.513* (0.276)

-0.335 (0.246)

-0.428 (0.410)

1.879*** (0.290)

2.089*** (0.258)

-0.837*** (0.219)

-1.301*** (0.156)

-1.765*** (0.331)

1.263

1.355

1.417

Adj. R2

0.993

0.991

0.991

LM t-stat

8.461***

6.862***

6.242***

14.887***

17.011***

14.887***

LM NR2 stat

32.175***

30.075***

27.075***

64.710***

63.540***

65.070***

-2.684*** [0.004] -1.981** [0.024]

-2.073** [0.019] -1.716** [0.043]

-2.319*** [0.010] -1.884** [0.030]

-4.256*** [0.000] -3.681*** [0.000]

-3.448*** [0.000] -3.086*** [0.001]

-4.361*** [0.000] -3.684*** [0.000]

B

B

B

B

β5 (POPDEN)t B

B

B

DW P

P

P

P

B

0.073 (0.074)

0.178 (0.149)

0.831*** (0.151)

0.823*** (0.227)

1.326

1.291

1.282

0.971

0.960

0.963

0.799*** (0.114)

Panel Unit Root Tests on Residuals LLC IPS

Notes: See notes to Table 2. The LM NR2 statistic is now compared to: χ2 (4) = 9.49 and χ2 (5) = 11.07. P

P

25

Table 4: FGLS Estimations of Local Unemployment: Relative Indices. ECM based on the Residuals of (7). ∆(Uit) = β0 + β1 ECMt-1 + β2 ∆(Unt) + β3 ∆(INDEXit) + β4 ∆(POPDENit) + β5 ∆(Uit-1) + νit (8) B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

(ECM)t-1

B

B

B

∆(U)nt

β2 B

B

β3 B

B

B

B

∆(RZI)t B

B

B

B

B

B

B

B

B

B

β6 ∆ (U)it-1 B

B

B

B

B

B

B

B

-0.479*** (0.090)

-0.608*** (0.075)

-0.446*** (0.041)

-0.429*** (0.048)

-0.446*** (0.041)

0.308** (0.137)

0.354*** (0.115)

0.320** (0.125)

0.996*** (0.067)

0.944*** (0.066)

0.974*** (0.075)

0.604* (0.344)

0.848* (0.500)

0.181 (0.445)

-0.006 (0.389)

8.922*** (1.652)

9.966*** (1.967)

8.437*** (1.819)

0.302*** (0.075)

0.255*** (0.094)

0.344*** (0.091)

B

β5 ∆ (POPDEN)t

B

-0.570*** (0.091)

∆(RDI)t

β4

B

6-LARGE CITIES PANEL

0.601* (0.349)

B

B

B

5-BORDER CITIES PANEL β1

B

0.662 (0.546)

0.615 (0.435)

0.304 (0.457)

-1.584 (0.959)

-1.466 (1.108)

-1.307 (1.076)

0.230*** (0.035)

0.228*** (0.052)

0.224*** (0.046)

Adj. R2

0.400

0.337

0.425

0.796

0.794

0.794

DW

2.096

2.082

2.135

2.150

2.167

2.190

LM t-stat

-0.606

-0.348

-0.087

-0.665

-0.908

-1.114

LM NR2 stat

1.820

3.120

1.235

0.858

0.936

1.794

P

P

P

P

Notes: The ECM term associated with β1 is the lagged one period residuals from (7) reported in the more general model of Table 3. The method of estimation is the FGLS-Cross Section SUR Weights. Below the coefficients are standard errors. In all specifications, the constant (β0) and fixed-effects that differ across states (β0i) are included but their estimates are not reported. The LM NR2 statistic is constructed as before and χ2 (6) = 12.59 for α = 5%. B

B

B

B

B

B

P

P