Jerarquía de las necesidades La jerarquía de necesida

The primary specification for the analyses in Section 1.3.4 is

∆ygkt=α+β(∆Igkt) + ΓXgkt+αgt+αkt+εgkt (1.3.1)

where g indexes a commuting zone, k indexes an industry group21, t indexes a year, and the ∆ operator represents a ten-year change within a commuting zone and industry group (change withingk). For example, in Columns 2 and 3 of Table 1.5, ∆ygktis the change in firm

presence in commuting zone g and industry group k during a given decade, divided by the start-of-decade workforce in gk. The independent variable of interest, ∆Igkt, is the change

in immigrant worker stock in gk between year t−10 and t, divided by the start-of-decade workforce in gk. In that specification, the coefficient of interest β measures the number of firms created, on net, per immigrant—the same interpretation as the results from Section 1.3.1. Xgkt is a vector of control variables that can include 1980 commuting zone-industry

characteristics interacted with year dummy variables and a Census region-industry-year fixed effect. The analyses covered by this specification span three decades (1980-2010), in which there were large immigrant inflows to the U.S., and covers nearly the entire geography of the U.S. (more than 700 commuting zones). Thus, it dramatically expands on the external validity and policy relevance of the results found in Section 1.3.1.

Even with the rich fixed effect structure contained in Equation (1.3.1), however, endogeneity concerns regarding immigrant industry choices within geographies and geographical choices within industry remain. Immigrant employees, for example, may choose to work in booming industries, generating biased estimates of β. Immigrant employees may also be more adept than native workers at locating to areas that are booming, as found in Cadena and Kovak (2016), even if they work in the same industry. Meanwhile, if immigrant entrepreneurs are attracted to geographies where they face less competition or if immigrant employees are linked to large firms in more concentrated markets, ordinary least squares (OLS) estimates of β with firm presence as an outcome may contain a downward bias. Measurement error may also play a role here, even with our sample restrictions, given that immigrant inflows can be small (and thus estimated from relatively few unweighted sample observations) within commuting zone-industry group and that Equation (1.3.1) is a panel model. This could generate substantial attenuation bias in any OLS estimate ofβ. In short, even in a relatively saturated model, isolating exogenous variation that pushes immigrants into commuting zone-industry pairings substantially strengthens causal interpretation of β at the cost of

21_{When noted, k may index a 1-digit SIC Sector. However, when not noted, it references the industry}

reducing estimate precision.

A standard shift-share instrument for immigrant stock Igt in a panel with commuting

zone-year observations would take the following form:

zgt ≡ 1 Eg,1980 X o πgo,1980×Io(−g)t

where 1980 serves as the base year and o indexes a worker’s origin country. πgo,1980 is the share of origin country o’s stock of immigrants in 1980 that was located in commuting zone

g, Io(−g)t is the overall stock of immigrants from countryo in year tfor all commuting zones

other thang, andEg,1980 is the 1980 workforce in commuting zone g. ∆zgt would then serve

as an instrument for ∆Igt, the overall change in immigrant presence in commuting zone

g between t −10 and t, by distributing immigrant inflows to commuting zones based on network effects operating through initial stocks πgo,1980.

Previous literature has recognized that replacing ∆Io(−g)t with variables that capture

exogenous push factors from sending country o in ∆zgt can make the exclusion restriction

more plausible. This notion accords with recent work by Borusyak et al. (2018), who demonstrate that when shift components are as good as randomly assigned conditional on shares, shift-share instruments do not violate the exclusion restriction. Llull (2017) thus uses conflict, natural disasters, changes in per capita income, and changes to political regimes as aggregate push factors. He takes the additional step of replacing πgo,1980 with distance because he works in a setting with cross-country migrant destinations.22 _{Monras (2015) hones} in on one sending country, Mexico, and uses the Peso Crisis of 1995 as an exogenous push factor, interacting it with 1980 state shares of Mexican immigrants to generate exogenous variation in Mexican inflows. Angrist and Kugler (2003) interact distance from Bosnia and Kosovo with indicators for years in which wars were taking place in those locations.

In this chapter, I follow a modified version of this strategy that also borrows from Autor et al. (2013). I take advantage of the German Institute for Employment Research (IAB) Brain-Drain Data, a unique source that contains counts of bilateral emigrant stocks from more than 150 sending countries worldwide residing in OECD member nations (Br¨ucker et al., 2013). I am thus able to replace Io(−g)t with Motnon-US, a measure of emigrants from

origin country o living in all OECD member nations other than the United States:

z_gtEmigrants = 1

Eg,1980

X

o

πgo,1980×Motnon-US

22_{That is, distance serves as an appropriate time-invariant, destination-specific interactor for Llull (2017)}

because it creates enough variation in how much it pushes sending country emigrants to different countries or different broad regions in North America. Distances between U.S. commuting zones and sending countries

∆z_gtEmigrants generates plausibly exogenous variation in immigrant locational decisions within the U.S. under the assumption that outflows from origin countries to non-U.S. OECD countries are unrelated to local economic outcomes in the U.S. (Borusyak et al., 2018), or if base-year sharesπgo,1980do not affect local economies during the study period (Goldsmith-Pinkham et al., 2018). Just as Autor et al. (2013) claim that Chinese exports to non-U.S. countries reflect increases in Chinese export productivity rather than product demand in the U.S., I claim that these outflows are much more likely to reflect migration push factors in sending countries rather than labor demand in the U.S.

Because Equation (1.3.1) is identified from within commuting zone-year variation, it requires an instrument that varies at the industry level as well. The standard shift-share approach accommodates this disaggregation by constructing:

z_gktStandard≡ 1

Egk,1980

X

o

πgo,1980×I(−g)okt

Here, the initial shares remain the same, but the aggregate component I(−g)okt is now the

number of immigrants from origin country o working in sector k in all commuting zones other than g at timet. Note that

I(−g)okt ≡[ρ(−g)okt×Io(−g)t]

where ρ(−g)okt is the proportion of origin o immigrants working in sector k in all commuting

zones other than the commuting zone of interest, g. Thus, it still utilizes the network effects provided byπgo,1980 for relevance, but does so separately by sector. While the IAB data does not separate emigrant outflows by sector, the restricted-access demographic data from the Census Bureau does contain detailed information on both the country of origin and industry of workers. Thus, in order to turn z_gtEmigrants into a geography-industry level instrument, I make the following adjustment:

z_gktEmigrants = 1

Egk,1980

X

o

πgo,1980×ρ(−r)okt×Motnon-US (1.3.2)

where r is the Census Region that commuting zone g resides in. This strategy takes advantage of the fact that immigrants from particular countries tend to specialize in certain industries due to comparative advantage, separate from demand in a particular local industry. Ultimately, z_gktEmigrantsthen predicts the number of immigrants residing in a given commuting zonegbased on network-induced locational preference, working in industrykdue to country-specific comparative advantage, and pushed into the U.S. by factors stemming from their country of

origin.

With our instrument fully detailed, we can now fully describe the utility of each fixed effect in Equation (1.3.1). For the decade ending in yeart,αgt removes any effects immigrant

inflows have at the commuting zone level as a whole. Under the premise that immigrants do not solely demand goods in the industry in which they work, αgt insulatesβ from being

identified by changes in consumption patterns that can result from immigration (e.g., Hong and McLaren, 2015). This premise is strengthened by the fact that we compare across 40 industry groups within a commuting zone—a level of detail allowed for by the granularity in each Census Bureau data source. Absent αgt, an inflow of immigrants into the “Hospitals”

industry group can generate an increase in economic activity in other nontradable industry groups because the new immigrant workers in the “Hospitals” industry group also consume goods and services locally. By including αgt and thus inducing comparison across industry

groups within a given commuting zone and decade, β measures the increase in economic activity in the “Hospitals” industry group above and beyond what other industry groups experienced due to this consumer demand effect.

αktplays an important role in ensuring instrument validity. ρ(−r)oktallocates immigrants

from origin country ointo industry groupk based on national level trends of industry choice for origin country o immigrants (excluding region r). In the absence of αkt, national level

shocks to industrykwould naturally allocate all workers towards industry groupk, regardless of origin. αkt precludes these shocks—which would induce both an increase in economic

activity in sector k across all commuting zones and in ∆z_gktEmigrants—from contaminating β. Instead, with the inclusion of αkt (or further, region-industry-year fixed effects), workers

from origin country o must be locating in industry group k above and beyond the national trend, and in regions where they are not affected by labor demand shocks in commuting zone

g. Thus, ∆z_gktEmigrants is more credibly sourced from immigrants’ comparative advantages in certain industries through ρ(−r)okt when αkt is included.

It is also useful to think through an example of how ∆z_gktEmigrantsprecludes reverse causality from identifying β. A simple and relevant example of labor demand pulling immigrants into specific geographies and industries comes from the housing bubble that metastasized between 2000 and 2005, largely in the South and West of the U.S. The housing bubble created a large labor demand shock for construction workers in the South and West Census Regions of the U.S., and induced immigrant workers from Mexico to fill this demand—the kind of inflow an instrumental variable should not use for identification ofβ. As seen in Panel A of Figure 1.4, Mexican inflows into the construction sector between 2000 and 2005 were more than 10 times larger than from the next closest country. As seen in Panel B of Figure 1.4, general immigrant inflows into the construction sector across commuting zones predominantly took

place in housing bubble cities throughout the South and West regions of the country. On the other hand, there is no reason to believe that the U.S. housing bubble would cause large outflows of Mexican emigrants to non-U.S. OECD countries. Furthermore, relative to the national trend, the change in propensity of Mexican immigrants to locate in the construction sector, outside of the South and West regions is not unusually strong. Thus, the aggregate component of ∆zEmigrants_gkt between 2000 and 2005

ρ(−r)ok,2005×Mo,non-US2005 −ρ(−r)ok,2000 ×Mo,non-US2000

should not reflect the labor demand shocks in construction that were occurring in the South and West of the country at that time. These two factors are illustrated in Panel A of Figure 1.5, where Mexico has a much more modest aggregate component for the construction sector between 2000 and 2005. The ultimate result of these corrections can be seen in Panel B, where the instrument-predicted immigrant inflows are far less concentrated in bubble cities. Results presented below, along with a series of analyses in Section A.2 test several many aforementioned and additional considerations regarding the validity of ∆z_gktEmigrantsmore systematically, with a particular focus on recently-formalized concerns regarding shift-share instrumentation. Table 1.5 and Section A.2.4 address instrument validity using various pre-trends tests. Section A.2.2 compares ∆z_gktEmigrantsto ∆zStandard

gkt along with another plausibly

exogenous replacement for the “shift” component in a typical migration instrument—lagged birth rates in origin countries. It finds results from this third instrument, ∆zBirths

gkt , and

∆z_gktEmigrants are nearly identical.23 _{Section A.2.3 demonstrates that migrant outflows to} non-US OECD countries are a relevant predictor of immigration to the U.S. at the origin country level. Section A.2.5 addresses concerns regarding correlated outcomes across observations with similar “share” components that can undermine inference when using shift-share instrumentation (Adao et al., 2019). It finds no evidence of such a problem in my triple-difference, commuting

zone-sector level regressions. Section A.2.6, along with Column 6 of Table 1.5, alleviates concerns about serial correlation in the “shift” component of the instrument (Jaeger et al., 2018). All told, the emigrants-based instrument passes a battery of tests meant to vet its validity for use. ∆zEmigrants_gkt and z_gktEmigrants therefore become the instruments of choice for

23_∆zBirths

gkt utilizes lagged birth counts in sending countries of individuals who are of prime migration age

by the time of our study period as a push factor. Because these are predetermined relative to the analysis,

they can be considered a better source of exogenous variation. The reason this chapter does not use ∆zBirths

gkt

as its preferred instrument, then, is out of practical rather than validity concerns. The demographic changes represented in ∆zBirths

gkt take years to unfold, making them much more amenable to decade (or longer horizon)

level analyses. Meanwhile, the analyses in Section 1.4 require an instrument at a five-year frequency. Thus, Section A.2.2 shows that ∆zBirths_gkt delivers near-identical results to ∆z_gktEmigrants. With this check in hand, I proceed with the emigrants instrument for the rest of the chapter.

Sections 1.3.4, 1.3.5 and 1.4.

1.3.4 Immigrant Workers and Firm Presence: a More Complete Accounting