Essays in Strategy and Microeconomics

CitationPoliquin, Christopher W. 2018. Essays in Strategy and Microeconomics. Doctoral dissertation, Harvard Business School.

Permanent linkhttp://nrs.harvard.edu/urn-3:HUL.InstRepos:41940970

Terms of UseThis article was downloaded from Harvard University’s DASH repository, and is made available under the terms and conditions applicable to Other Posted Material, as set forth at http://nrs.harvard.edu/urn-3:HUL.InstRepos:dash.current.terms-of-use#LAA

Share Your StoryThe Harvard community has made this article openly available.Please share how this access benefits you. Submit a story .

Accessibility

Essays in Strategy and Microeconomics

A dissertation presentedby

Christopher W. Poliquinto

The Strategy Unit at Harvard Business School

in partial fulfillment of the requirementsfor the degree of

Doctor of Business Administrationin the subject of

Business Administration

Harvard UniversityCambridge, Massachusetts

March 2018

©2018 – Christopher W. Poliquinall rights reserved.

Dissertation Advisor: Professor Shane Greenstein Christopher W. Poliquin

Essays in Strategy and Microeconomics

Abstract

This dissertation consists of three essays.

In Chapter 1, I study the beneficiaries of technology adoption in the workplace. I

combine worker-level wage data with information on broadband adoption by Brazilian

firms to estimate the effects of broadband on wages. Overall, wages increase 2.3 per-

cent following broadband adoption. Consistent with the theory of biased technological

change, wages increase the most for workers engaged in non-routine cognitive tasks and

returns are negative for routine cognitive tasks. There is no effect of broadband adop-

tion on wages for either routine or non-routine manual tasks. Additionally, I estimate

the effect of broadband on selected quantiles of the within-firm wage distribution and

find evidence that within-firm wage inequality increases following broadband adoption.

Both new hires and the firm’s existing employees benefit from broadband adoption,

which indicates that broadband’s effects are not driven only by better recruitment of

new employees.

Chapter 2 presents three main findings about the impact of mass shootings on gun

policy in the United States. First, mass shootings evoke large policy responses. A

single mass shooting leads to a 15 percent increase in the number of firearm-related bills

iii

Dissertation Advisor: Professor Shane Greenstein Christopher W. Poliquin

introduced within a state in the following year. This effect increases with the number

of fatalities. Second, mass shootings account for a small portion of all gun deaths, but

have an outsized influence relative to other homicides. Our estimates suggest that the

per-death impact of mass shootings on bills introduced is about 80 times as large as the

impact of individual gun homicides in non-mass shooting incidents. Third, when looking

at enacted laws, the impact of mass shootings depends on the political party in power.

A mass shooting increases the number of enacted laws that loosen gun restrictions by

75 percent in states with Republican-controlled legislatures. There is no statistically

significant effect of mass shootings on laws enacted when there is a Democrat-controlled

legislature.

Chapter 3 directly studies the extent and drivers of internal labor markets in multi-

business firms. Leveraging a rich employer-employee matched dataset from Brazil, we

track all worker movements across firm units. We find that multi-business firms source

a large share of their workers internally, especially managers and workers with more

firm-specific experience. Redeployed workers earn a large wage premium over otherwise

comparable workers hired through external labor markets. Geographic proximity and

resource relatedness between establishments play an important role in facilitating rede-

ployment. In contrast to prevailing views of internal labor markets as a means to avoid

external labor market frictions, our findings are consistent with internal labor markets

as conduits of knowledge.

iv

Contents

1 The Effect of the Internet on Wages 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Anecdotal Evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.4 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2 The Impact of Mass Shootings on Gun Policy 282.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282.2 Background and Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3 Internal Labor Markets in Multi-business Firms 523.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.2 Theory and Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . 593.3 Data and Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . 713.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

Appendix A Construction of O∗NET Task Measures 89

Appendix B Variable Definitions for Chapter 2 92

Appendix C Coding Gun Laws 94

Appendix D Effect of Mass Shootings in Neighboring States 96

Appendix E Effect of Mass Shootings on Enacted Laws 100

v

Appendix F Predicting Mass Shootings 102

Appendix G Mass Shootings and State-Specific Time Trends 106

Appendix H Placebo Mass Shooting Analyses 108

Appendix I Excluding States from Mass Shooting Analyses 110

Appendix J Gun Ownership, Shootings, and Enacted Laws 114

Appendix K Mass Shootings as an Instrument for Gun Policy 116

References 125

vi

Acknowledgments

I am grateful to Shane Greenstein, Michael Luca, Raffaella Sadun, and Deepak Mal-

hotra for their advice and guidance. I also thank the HBS Latin America Research

Center, Partners of the Americas, Harvard Business School, the Harvard Economics

Department, and Brazil’s Ministério do Trabalho e Emprego for supporting this project.

I also acknowledge the immense support of my friends and family.

Chapter 2 is co-authored with Michael Luca and Deepak Malhotra. We thank Joseph

Hall and Jessica Li for excellent research assistance.

Chapter 3 is co-authored with Jasmina Chauvin.

vii

1The Effect of the Internet on Wages

1.1 Introduction

Who benefits from technology adoption in the workplace? Technology can substitute

for some workers while complementing others. Specifically, the “task approach” to

labor markets highlights the potential for digital technologies to substitute for workers

in performing routine tasks, while complementing workers in non-routine tasks (Autor,

Levy, and Murnane, 2003). To date, empirical work on this hypothesis has largely relied

on industry-, region-, or to a lesser extent, firm-level data. In contrast, this paper uses

worker-level wage data in conjunction with firm-level information on technology use

over time. Specifically, I study how broadband Internet technology affects the wages of

individual workers within firms.

1

I find that wages increase 2.3 percent following firm broadband adoption, but the

effect of broadband is heterogeneous. Regressions of wages on the task profile of jobs

suggest that broadband complements employees performing non-routine cognitive tasks,

while substituting for workers in routine cognitive tasks. Intuitively, both routine and

non-routine manual tasks are unaffected by broadband.

The differences in the returns to broadband across tasks have implications for within-

firm wage inequality. I examine changes to the entire wage distribution within firms fol-

lowing broadband adoption using a grouped quantile regression estimator (Chetverikov,

Larsen, and Palmer, 2016). Wage increases following broadband are concentrated in the

right tail of the wage distribution; in other words, within-firm wage inequality increases

after broadband adoption. This result contributes to a literature that emphasizes the

role of firms in determining pay inequality (e.g. Cobb, 2016; Gartenberg and Wulf,

2017b; Nickerson and Zenger, 2008), and provides the first direct evidence connecting

adoption and use of advanced information technology to a widening pay gap within an

organization.

As evidence of broadband enhancing the productivity of existing workers, rather

than only improving the recruitment of new workers, I show that wages increase for

both new hires and existing employees following broadband adoption. Furthermore,

firm directors—who are most likely to also be firm owners—appear to capture large

rents from the introduction of broadband, a pattern consistent with increased firm

productivity post-adoption.

2

The analysis combines an employer-employee matched dataset from Brazil with firm-

level data on technology use over time. By linking information on which firms use

broadband with data on their individual workers, I can estimate the effect of broadband

within firms over time. Additionally, I can examine changes in the wages of individual

workers while controlling for worker characteristics and unobserved firm heterogeneity.

This paper is the first to combine within-firm variation on technology use with large-

sample microdata on the wages and characteristics of individual workers. While other

research has examined the impact of technology —including the Internet—on wages,

prior studies have not observed changes in the technology used at individual firms over

time. Recent research on the effects of the Internet in Brazil (Almeida, Corseuil, and

Poole, 2017; Dutz et al., 2017), Africa (Hjort and Poulsen, 2017), Norway (Akerman,

Gaarder, and Mogstad, 2015), and the United States (Forman, Goldfarb, and Green-

stein, 2012; Gillett et al., 2006; Kolko, 2012) relies on geographic variation in Internet

availability and/or cross-sectional variation in firm adoption. In contrast, I observe the

same firm and workers before and after the adoption of broadband. The results of this

paper are consistent with prior work, which shows broadband substitutes for workers

engaged in routine tasks while complementing workers engaged in non-routine tasks.

Broadband technology is especially worthy of study because of the Internet’s perva-

siveness and policymakers’ interest in public investments in broadband infrastructure.

Nearly 50 percent of people worldwide now access the Internet. The transformation of

the Internet from a technology used by fewer than one percent of people in the mid-

3

1990s to the ubiquitous network of today has potentially large effects on firm operations

and jobs.

Although a number of studies suggest that broadband, and Internet access generally,

is a skill-biased technological change, few if any provide concrete examples of how or why

this might be the case. The next section provides anecdotal evidence from interviews

with Brazilian managers suggesting that broadband use in firms can assist workers

with non-routine cognitive tasks while substituting for workers in performing routine

cognitive tasks.

1.2 Anecdotal Evidence

This section provides examples, through interviews with managers in Brazilian firms, of

how broadband can affect firm operations. Although several papers suggest broadband

complements workers in performing non-routine tasks, while substituting for routine

tasks, few are specific about how high-speed Internet access might do this.

Managers I interviewed described using broadband to facilitate information exchange

both within and between firms and customers. A manufacturer of industrial equipment

explained how broadband provided constant connectivity with their suppliers that al-

lowed them to automate routine aspects of inventory management:

“We scan the barcode on the kanban card and new part orders are sentdirectly to the supplier. This has saved time for the logistics people tospend more time on other tasks, like inventory optimization. It also means

4

we’ve had some layoffs. We need fewer people to do ordering, and a differentset of skills.”

The same firm also used broadband to facilitate communication between workers

directly involved in production and engineers and managers higher in the organizational

hierarchy. Broadband, therefore, complemented the skills of engineers in the non-routine

task of reviewing product design issues and communicating solutions:

“The [machine] operator scans the production order and the computer down-loads the CAD drawing from our database. We can share designs worldwide.If there is a problem, he can hit a button on the screen and report it to anengineer, who can diagnose and solve it.”

A provider of medical imaging services explained using broadband to automate ap-

pointment scheduling, therefore eliminating the routine task of finding open dates. At

the same time, this firm leveraged broadband to unify databases across multiple work

sites in a single location so that important documents could be shared and accessed

from anywhere. This made it easier for doctors to access patient medical records across

facilities.

A manufacturer of bottled water used broadband to connect its machines to the

company that supplied them so that their performance could be monitored remotely.

This change obviated the need for someone who could monitor the machine’s controls,

eliminating the routine task of documenting and recording information.

In addition to these examples, managers reported using broadband to stay in closer

5

contact with their customers, research competitors, and communicate with subordi-

nates.

Two aspects of these examples are especially noteworthy. First, the examples em-

phasize that broadband is a tool that firms use in conjunction with other software and

hardware to enable changes in production. Second, the examples illustrate that work-

ers can benefit from the introduction of broadband without using it in their own work.

The changes in inventory management described above benefited workers with skills in

optimization even though broadband was being used to facilitate part orders, not solve

optimization problems. The wage effects reported in this paper are not the treatment

effects of assigning broadband to particular workers. Rather, the analysis examines

the impact of firm broadband adoption and concomitant changes in production on the

wages of all the firm’s workers.

1.3 Data

The data used in this study are richer than data used in prior studies of broadband

adoption because they include information on individual workers and their employers

over time. This allows me to examine how wages change for different types of workers

following firm adoption of broadband.

Data on individual workers come from the Relação Anual de Informações Sociais

(RAIS) for the years 2000 to 2009. RAIS is an establishment-employee matched survey

6

of all employers in Brazil’s formal economy conducted annually by the Ministério do

Trabalho e Emprego (MTE). Participation is mandatory. Unique identifiers for workers

and establishments in RAIS allow records to be linked across years. Employee records

include data on wages, occupation, education, experience, age, gender, and contract

hours (but not hours actually worked).

I combine the employer-employee matched data from RAIS with firm-level data on

broadband adoption from the Latin American version of the Ci Technology Database

(CiTDB) from Aberdeen Group.1 The European and U.S. versions of CiTDB have been

used in prior studies to measure technology adoption (Bloom et al., 2014). CiTDB

contains information on communication technologies used by the firm (e.g. xDSL, T1,

etc.), which I use to measure broadband adoption.

I limit my study to manufacturing firms—which is the largest group of businesses in

the data—with technology adoption information in Harte Hanks and wages in RAIS so

that analyses of the task content of jobs and occupational hierarchy can be more easily

interpreted.

Figure 1.1 shows that broadband use increased substantially from 2000 to 2009; fewer

than 20 percent of the sample firms used broadband in 2000, but more than 70 percent

had a broadband connection by 2009. Note that these numbers are not necessarily repre-

sentative of all Brazilian manufacturing firms. The firms surveyed by Harte Hanks—my

1CiTDB and Aberdeen Group were formerly owned by Harte Hanks; Halyard Capital acquiredAberdeen and CiTDB in April 2015.

7

Figure 1.1: Adoption of High-Speed Internet

010

2030

4050

6070

80

Firm

s w

ith H

igh-

Spe

ed In

tern

et (

%)

2000 2003 2006 2009Year

source of technology data—are larger than the typical firm in Brazil.

To examine how the effects of broadband vary for different types of workers, I use

measures from the U.S. Department of Labor’s O∗NET database to characterize the

importance of various tasks for each occupation.2 O∗NET contains hundreds of scales

that rate the importance of various activities, skills, abilities, and work contexts for

each job. For consistency with prior research and to limit researcher degrees of freedom

in picking from hundreds of O∗NET scales (Autor, 2013), I use the same variables as

Acemoglu and Autor (2011) and computer code from David Autor’s website3 to produce

2I use O∗NET version 9.0, which was released in December 2005 and is the most recent versionto use the SOC 2000 occupation codes. I use this version because I rely on a crosswalk betweenSOC 2000 and ISCO 88 to match the Brazilian occupation codes with O∗NET.

3Available at https://economics.mit.edu/faculty/dautor/data/acemoglu (archived at

8

Table 1.1: Task Summary Statistics

Task measure mean sd p5 p10 p50 p90 p95

Non-routine cognitive -0.72 0.79 -1.61 -1.60 -0.85 0.30 0.70Non-routine manual 0.32 0.74 -0.94 -0.90 0.39 1.30 1.40Routine cognitive -0.19 0.77 -1.36 -1.02 -0.43 1.13 1.17Routine manual 0.79 0.99 -0.68 -0.46 0.59 2.16 2.16

Note: Table shows the distribution of occupation task measures acrossindividual workers.

four measures of the extent to which each occupation involves various tasks:

1. Non-routine cognitive

2. Non-routine manual

3. Routine cognitive

4. Routine manual

Each of these variables is standardized across occupations so that a unit increase

equals a one standard deviation increase in the extent to which an occupation depends

on the given tasks relative to other occupations. Appendix A lists the specific O∗NET

scales used for each task measure. Table 1.1 shows the distribution of the task measures

across Brazilian workers. The means and medians for the cognitive (manual) scales are

negative (positive), reflecting the greater prevalence of workers engaged in manual-task-

intensive occupations in Brazil’s manufacturing sector.

O∗NET scales were developed to measure features of U.S. occupations. I adapt

these measures to Brazil by merging both the U.S. and Brazilian occupation codes

https://perma.cc/B7SK-VKUV).

9

Table 1.2: Wage Distribution by Hierarchy Level

Director Manager Supervisor Worker

mean 18,085 8,679 3,763 1,476p5 1,593 1,053 735 391p10 3,030 1,692 984 468p25 7,573 3,458 1,674 636p50 16,617 7,144 2,953 961p75 26,403 11,767 4,979 1,648p90 35,531 17,166 7,365 2,937p95 40,745 21,779 9,145 4,222

Note: Wages are mean monthly wage in 2008 reais.

to the International Standard Classification of Occupations (ISCO 88). This results in

instances where a single Brazilian occupation code matches multiple U.S. codes; in these

cases I assign the Brazilian occupation to a simple average of the U.S. task measures.

Additionally, I use occupation codes from RAIS to divide each establishment’s work-

force into hierarchical layers. My approach mirrors the method used by Caliendo, Monte,

and Rossi-Hansberg (2015) in their study of French manufacturers. Specifically, I assign

each worker to one of the following four layers:

1. Directors (e.g. Chief Executive Officer, Chief Financial Officer)

2. Managers (e.g. Sales Manager, Branch Manager)

3. Supervisor (e.g. Foreman, Logistics Supervisor)

4. Workers (e.g. Welder, Production Line Feeder, Fish Cooker)

Like Caliendo, Monte, and Rossi-Hansberg (2015), I find the grouping of occupations

into layers reflects meaningful differences between employees. Table 1.2 shows the mean

10

and selected percentiles of the wage distribution by layer. Directors and managers have

higher wages than supervisors, who have higher wages than workers (at all percentiles).

1.4 Methodology

I use a staggered difference-in-differences research design that identifies the effect of

broadband adoption on wages by comparing firms that did and did not adopt broadband

over the ten-year period between 2000 and 2009.

The main models of interest examine the effect of broadband adoption on workers,

allowing for the effect of broadband to differ by occupation:

lnwijt = β0Djt + β′1Djt ∗Kit + θ′Kit + δ′Xijt + γLjt + αj + λκ(j)t + ϵijt (1.1)

where wijht is the real wage of worker i at firm j in year t. Djt is an indicator variable

for broadband use by firm j and Kit is a vector of continuous measures representing the

task content of worker i’s occupation in year t. The task measures capture the extent to

which the worker’s job involves routine vs. non-routine and cognitive vs. manual tasks.

The vector Xijt is a set of time-varying worker covariates that includes education, current

job experience, sex, age, age squared, and log contract hours.4 Some specifications also

include log employment, Ljt, to control for the possibility that larger firms pay higher

wages and are more likely to adopt broadband (Oi and Idson, 1999). Employment,

4The data do not include actual hours worked, but do include hours in the labor contract.Full-time work in Brazil is 44 hours per week.

11

however, could itself be affected by broadband adoption; I therefore use employment

as the dependent variable in other analyses and omit it from most models. The model

includes both firm (αj) and industry-year (λκ(j)t, where κ(j) is the industry of firm

j) fixed effects that control for unobserved firm heterogeneity and annual shocks that

affect all workers within an industry equally.

Combining employer-employee matched data with information on technology use over

time allows me to examine how the entire wage distribution within firms changes fol-

lowing broadband adoption. To do so, I implement the grouped quantile regression

approach from Chetverikov, Larsen, and Palmer (2016). Specifically, I estimate:

Qlnwijt|Djt,ηjt(τ) = αj(τ) + λκ(j)t(τ) + γ′(τ)zij + β(τ)Djt + ϵ(τ, ηjt) (1.2)

where Q(τ) selects the τth quantile of log wages for firm j in year t, Djt is an indicator

for firm broadband adoption, zij is a vector of individual-level covariates, and αj and

λκ(j)t are firm and industry-year fixed effects.

The grouped quantile approach allows me to estimate how broadband adoption affects

inequality within firms. Greater effects of broadband in the upper quantiles of the wage

distribution than in lower quantiles imply that inequality within firms increases following

broadband adoption.

In addition to studying the effect of broadband on wages, I also examine how employ-

12

Table 1.3: Summary Statistics

mean sd p5 p10 p50 p90 p95

High-speed Internet 0.52 0.50 0 0 1 1 1Log wage 7.05 0.81 6.0 6.2 6.9 8.2 8.6Log contract hours 3.77 0.09 3.7 3.7 3.8 3.8 3.8Tenure in months 60.45 70.92 1.9 3.4 32.7 161.9 211.9Age 33.11 10.10 20.0 21.0 32.0 47.0 52.0Female 0.24 0.43 0 0 0 1 1Education DummiesBelow Elementary 0.08 0.27 0 0 0 0 1Elementary 0.09 0.28 0 0 0 0 1Some Middle School 0.14 0.35 0 0 0 1 1Middle School 0.15 0.35 0 0 0 1 1Some High School 0.10 0.31 0 0 0 1 1High School 0.31 0.46 0 0 0 1 1Some College 0.05 0.21 0 0 0 0 0Higher Ed Degree 0.08 0.28 0 0 0 0 1

Note: Log wages are log of mean monthly wage in 2008 reais.

ment changes at the firm level following broadband adoption:

Ljt = βDjt + αj + λκ(j)t + ϵjt (1.3)

Table 1.3 presents summary statistics of variables used in the analyses. Just over half

of observations are for people working in firms that use broadband.

1.5 Results

1.5.1 Wage Effects

Overall, wages increase 2.3 percent following firm adoption of broadband. Table 1.4

shows the effect of broadband adoption without distinguishing between occupations or

13

Table 1.4: Wage Effects of Broadband

(1) (2) (3) (4) (5)

Broadband 0.034∗∗∗ 0.026∗∗∗ 0.026∗∗∗ 0.023∗∗∗ 0.023∗∗∗(0.009) (0.009) (0.009) (0.008) (0.008)

Log employees 0.015∗∗ -0.007(0.007) (0.008)

Worker Controls • • • •Fixed EffectsFirm • • • • •Year • • •Industry-Year • •

Adj-R2 0.45 0.69 0.69 0.69 0.69Firms 3,333 3,333 3,333 3,332 3,332N 6,949,890 6,949,890 6,949,890 6,949,887 6,949,887

Note: Standard errors in parentheses are clustered by firm.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

types of employees. The results in columns 2–3 include firm and year fixed effects,

while columns 4–5 include firm and industry-year fixed effects. The estimates are stable

across specifications and show a positive average effect of broadband adoption on wages.

Comparing the results of columns 2 and 4 with those of columns 3 and 5 shows that the

estimate of the broadband effect is insensitive to controlling for firm size. The increase

in wages following broadband adoption, therefore, is not explained by bigger, growing

firms paying both higher wages and simultaneously choosing to adopt broadband.

There are several caveats to a causal interpretation of these results. First, firms might

increase wages for other reasons that happen to coincide with broadband adoption.

Without controlling for these factors, wage increases will be erroneously attributed to

broadband. Second, even if broadband causes wages to increase, the firms most likely

14

to benefit from the technology will be more likely to adopt, in which case estimates

from the sample of adopters will be greater than the effect of introducing other firms

to broadband. Third, trends in wages prior to broadband adoption might be different

from trends in wages at firms that do not adopt. In this case, firms that do not adopt

broadband are a poor control group for the adopters.

I cannot correct for omitted variables without an instrument. The problem of firms

selecting into broadband use, however, is partially mitigated by the ten-year sample

period. Figure 1.1 shows that most firms in the sample eventually adopt broadband.

Additionally, the long sample period allows me to examine wage trends prior to broad-

band adoption. Figure 1.2 shows coefficient estimates from a modified version of the

model in column 4 of Table 1.4 that includes separate dummy variables for years before

and after adoption. These single year estimates are imprecise, but show that the largest

wage increases happen in the years following broadband adoption. There is, however,

some evidence that wages at adopting firms begin increasing relative to non-adopting

firms in the year before broadband adoption.

The effect of broadband is heterogeneous; workers in occupations that require more

non-routine cognitive tasks see larger wage gains than workers in occupations that are

intensive in routine cognitive tasks. Table 1.5 shows regressions in which broadband

adoption is interacted with occupation-specific measures of task intensity. The coeffi-

cients on non-routine cognitive and routine cognitive tasks have opposite signs, suggest-

ing that broadband complements workers performing non-routine cognitive tasks and

15

Figure 1.2: Wages Before and After Broadband Adoption

-0.02

0.00

0.02

0.04

0.06

0.08

Coe

ffici

ent E

stim

ate

3 YearsPrior

2 YearsPrior

1 YearPrior

AdoptionYear

1 YearAfter

2 YearsAfter

3+ YearsAfter

Note: Values along the x-axis represent time relative to broadband adoption; e.g. “2 YearsAfter” refers to the second year following adoption.

16

substitutes for workers in routine cognitive tasks. A one unit increase (roughly one

standard deviation) in the intensity of non-routine cognitive tasks implies an additional

3.5–4.5 percent wage increase following broadband adoption. In contrast, a one unit in-

crease in the intensity of routine cognitive tasks implies a 4–5 percent decrease in wages,

which cancels out the baseline increase of four percent from broadband adoption.

Table 1.5 also indicates that the effect of broadband adoption does not vary in the

intensity of manual tasks. This is consistent with the intuition that broadband ought

to have small, if any, effect on tasks that require interaction with equipment and using

one’s hands.

The use of four, continuous task measures interacted with broadband complicates

interpretation of the results in Table 1.5. Table 1.6 therefore presents the distribution

of wage effects (across workers) implied by the task regressions. For each worker, I

use the coefficients from the regressions in Table 1.5 and the task intensities of the

worker’s occupation to calculate the hypothetical impact of broadband for that worker.

I then calculate the distribution of these wage effects across all workers. The results in

Table 1.6 show that the effect of broadband on real wages is positive for the majority

of workers and that wage gains in the right tail of the distribution are much larger in

magnitude than wage losses in the left tail.

Overall, the broadband/task interaction effects of Table 1.5 are consistent with the

routinization hypothesis (Autor, Levy, and Murnane, 2003) that computer technology

complements and increases demand for non-routine tasks while substituting for routine

17

Table 1.5: Wage Effects of Broadband by Tasks

(1) (2)

Broadband ×Intercept 0.041∗∗∗ 0.040∗∗∗

(0.011) (0.012)

Non-routine cognitive 0.044∗∗ 0.036∗(0.018) (0.019)

Non-routine manual -0.004 -0.009(0.009) (0.009)

Routine cognitive -0.050∗∗∗ -0.041∗∗(0.016) (0.018)

Routine manual 0.006 0.006(0.008) (0.008)

Non-routine cognitive 0.145∗∗∗ 0.150∗∗∗(0.017) (0.017)

Non-routine manual -0.059∗∗∗ -0.056∗∗∗(0.007) (0.007)

Routine cognitive -0.018 -0.023(0.016) (0.016)

Routine manual -0.006 -0.007(0.007) (0.007)

Worker Controls • •Fixed EffectsFirm • •Year •Industry-Year •

Adj-R2 0.70 0.71Firms 3,333 3,332N 6,871,375 6,871,372

Note: Standard errors in parentheses are clustered by firm.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

18

Table 1.6: Summary of Wage Effects of Broadband by Tasks

Model mean sd p5 p10 p25 p50 p75 p90 p95

(1) 0.023 0.036 -0.020 -0.015 0.005 0.015 0.035 0.079 0.115(2) 0.024 0.029 -0.013 -0.000 0.010 0.019 0.033 0.072 0.096

Note: Table shows the distribution of wage effects across individual workers forthe models in Table 1.5.

tasks. In the case of broadband, this pattern is pronounced for cognitive tasks, but not

present for manual tasks.

My estimates for the wage effects of broadband are larger, although roughly similar

in magnitude, to those of Dutz et al. (2017), who examine the regional wage effects

of Brazil’s Internet (but not specifically broadband) rollout. They report a two-year

cumulative wage increase of 4.1–4.8 percent for middle- and high-skill occupations in

manufacturing in response to an increase in Internet access, but no wage effect for low-

skill occupations.5 A possible explanation for the larger effect estimates in this paper

is that, unlike Dutz et al. (2017), I observe the adoption decisions of individual firms

instead of relying on measures of regional broadband availability.

1.5.2 New Versus Existing Employees

The effect of broadband on new employees is the same as the effect on existing employees.

This suggests that wage increases from broadband adoption are not driven only by

firms recruiting better workers post-adoption. Table 1.7 shows the effect of broadband

5Internet access in Dutz et al. (2017) is measured using the share of schools with Internet ineach municipality. The reported effects are based on increasing Internet access from 0 to 100percent (i.e. going from no access to every school having access).

19

adoption on wages allowing for the effect to differ by whether an employee is in his first,

first two, or first three years of working at the firm. The results show that newly hired

employees do not earn an additional wage premium from broadband adoption over that

earned by existing employees.

1.5.3 Wage Effects and Organizational Hierarchy

Wage increases following broadband adoption are greatest for workers higher in the

organizational hierarchy: directors and managers see larger increases than lower-level

workers. Columns 1 and 4 of Table 1.8 show that directors and managers earn 8–9

percent more following broadband adoption compared to a main effect of just over two

percent for all employees.

The effect of broadband is especially large for directors at the top of the organiza-

tional hierarchy. Columns 2–3 and 5–6 split the managers and directors group into

two separate coefficients, and columns 3 and 6 add another coefficient for supervisors,

who are grouped with workers in the other columns. The estimates suggest that di-

rectors earn 18–19 percent more following firm adoption of broadband. This is about

nine percentage points more than the increase for managers. Most firms in the sample

are private companies. The directors in this sample are therefore more likely to have

an ownership stake in the firm than if the firms were public. The wage increases for

directors are consistent with firm owners capturing large gains as a result of broadband

increasing firm productivity. Unfortunately, I do not have data on revenue or non-labor

20

Table 1.7: Wage Effects, New vs. Existing Employees

(1) (2) (3) (4) (5) (6)

Broadband ×Intercept 0.023∗∗∗ 0.023∗∗∗ 0.024∗∗∗ 0.021∗∗∗ 0.020∗∗ 0.022∗∗

(0.008) (0.008) (0.008) (0.008) (0.008) (0.009)

Hiring year 0.009 0.008(0.007) (0.007)

First 2 years 0.003 0.001(0.008) (0.007)

First 3 years 0.001 -0.002(0.008) (0.008)

Hiring year -0.136∗∗∗ -0.136∗∗∗(0.005) (0.005)

First 2 years -0.155∗∗∗ -0.154∗∗∗(0.005) (0.005)

First 3 years -0.152∗∗∗ -0.151∗∗∗(0.006) (0.005)

Worker Controls • • • • • •Fixed EffectsFirm • • • • • •Year • • •Industry-Year • • •

Adj-R2 0.69 0.69 0.69 0.69 0.70 0.70Firms 3,333 3,333 3,333 3,332 3,332 3,332N 6,949,890 6,949,890 6,949,890 6,949,887 6,949,887 6,949,887

Note: Standard errors in parentheses are clustered by firm.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

21

Table 1.8: Wage Effects of Broadband by Hierarchy Level

(1) (2) (3) (4) (5) (6)

Broadband ×Intercept 0.024∗∗∗ 0.024∗∗∗ 0.024∗∗∗ 0.022∗∗∗ 0.022∗∗∗ 0.023∗∗∗

(0.009) (0.009) (0.009) (0.008) (0.008) (0.008)

Director/Manager 0.052∗∗ 0.063∗∗∗(0.023) (0.021)

Director 0.141∗∗∗ 0.141∗∗∗ 0.153∗∗∗ 0.153∗∗∗(0.042) (0.043) (0.042) (0.043)

Manager 0.050∗∗ 0.051∗∗ 0.061∗∗∗ 0.061∗∗∗(0.023) (0.024) (0.021) (0.022)

Supervisor 0.002 0.004(0.014) (0.013)

Director/Manager 0.723∗∗∗ 0.718∗∗∗(0.021) (0.019)

Director 1.163∗∗∗ 1.230∗∗∗ 1.152∗∗∗ 1.220∗∗∗(0.033) (0.034) (0.033) (0.034)

Manager 0.688∗∗∗ 0.741∗∗∗ 0.683∗∗∗ 0.737∗∗∗(0.021) (0.022) (0.019) (0.019)

Supervisor 0.469∗∗∗ 0.467∗∗∗(0.010) (0.010)

Worker Controls • • • • • •Fixed EffectsFirm • • • • • •Year • • •Industry-Year • • •

Adj-R2 0.70 0.70 0.71 0.71 0.71 0.72Firms 3,333 3,333 3,333 3,332 3,332 3,332N 6,949,890 6,949,890 6,949,890 6,949,887 6,949,887 6,949,887

Note: Standard errors in parentheses are clustered by firm.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

22

inputs to explore this hypothesis. Akerman, Gaarder, and Mogstad (2015), however,

report that firms in Norway earn large rents from broadband adoption, and Jung and

López-Bazo (2017) find a positive effect of broadband on regional productivity in Brazil.

The greater effect of broadband for directors and managers implies that within firm

wage inequality increases following adoption. To more thoroughly examine this pattern,

I use the grouped quantile regression estimator from Chetverikov, Larsen, and Palmer

(2016) to assess how broadband adoption affects the distribution of wages within firms.

Figure 1.3 plots the effect of broadband on selected quantiles of the wage distribution.

Figure 1.3a shows results from a model without worker-level controls, while Figure 1.3b

is based on a model that controls for experience, age, and two education dummies (high

school and college completion). The sample in the latter model is also restricted to

firms with at least 10 employees to allow for the inclusion of the worker-level controls.

Although the estimates for the individual quantiles are imprecise, the pattern of point

estimates in Figure 1.3 suggests that broadband has larger effects on the right tail of

the wage distribution than on wages in the left tail. In other words, high wage workers

benefit more than low wage workers from broadband adoption and inequality within

firms increases.

Broadband’s effect in widening the within-firm wage distribution is noteworthy for

the literatures on vertical pay comparisons within firms (e.g. Gartenberg and Wulf,

2017a,b; Kacperczyk and Balachandran, 2018), the antecedents of compensation poli-

cies (e.g. Chin and Semadeni, 2017; Fredrickson, Davis-Blake, and Sanders, 2010), and

23

Figure 1.3: Quantile Effects of Broadband Adoption

-0.010

0.000

0.020

0.040

0.060

0.080B

road

band

Coe

ffici

ent

0 20 40 60 80 100

Percentile

(a) Without worker micro-covariates

-0.020

0.000

0.020

0.040

0.060

0.080

0.100

0.120

0.140

Bro

adba

nd C

oeffi

cien

t

0 20 40 60 80 100

Percentile

(b) With worker micro-covariates

24

the role of firms in determining pay inequality (e.g. Cobb, 2016). This paper provides

the first direct evidence connecting adoption and use of advanced information technol-

ogy to a widening pay gap within an organization. Furthermore, this paper provides

estimates of broadband’s effect across the entire wage distribution; existing research on

pay dispersion is largely focused on top-management teams and key employees.

Prior work suggests that pay inequality can have psychological costs (Larkin, Pierce,

and Gino, 2012), and can negatively impact performance (Fredrickson, Davis-Blake,

and Sanders, 2010; Siegel and Hambrick, 2005). Unfortunately, I do not have data to

assess either the first order effect of broadband on performance or any second order

effects operating through employee motivation. I also lack data on performance-linked

compensation that would allow me to examine how different components of pay change

in response to technology adoption.

1.5.4 Employment Effects

Broadband has positive effects on firm-level employment. Column 1 of Table 1.9 in-

dicates that employment increases roughly 5.4 percent following broadband adoption.

Columns 2 and 3 show separate regressions for managerial and non-managerial employ-

ees respectively. These estimates are not statistically different from zero at conventional

significance levels, and the point estimates do not suggest different employment effects

for workers and managers following broadband adoption. Columns 4–6, which include

industry-year fixed effects instead of just year fixed effects, show slightly larger esti-

25

Table 1.9: Employment Effects of Broadband

Total Managers Workers Total Managers Workers(1) (2) (3) (4) (5) (6)

Broadband 0.053∗∗ 0.044∗ 0.040 0.071∗∗∗ 0.054∗∗ 0.058∗∗(0.026) (0.026) (0.027) (0.027) (0.026) (0.026)

Worker Controls • • • • • •Fixed EffectsFirm • • • • • •Year • • •Industry-Year • • •

Adj-R2 0.84 0.79 0.85 0.84 0.81 0.76Firms 3,026 2,722 3,023 2,990 2,990 2,990N 17,722 15,348 17,696 17,310 17,310 17,310

Note: Standard errors in parentheses are clustered by firm.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

mates. Column 4 indicates that employment increases about seven percent following

broadband adoption, and columns 5–6 again suggest that the effect is similar for man-

ages and non-managers.

1.6 Conclusion

I combine data on firm adoption of broadband technology over time with data on indi-

vidual workers to estimate the effects of broadband on wages and employment. Overall,

wages increase 2.3 percent following broadband adoption, but the effects are heteroge-

neous. Consistent with the theory of skill-biased technological change, wages increase

the most for workers engaged in non-routine cognitive tasks. Returns for routine cog-

nitive tasks are negative, and intuitively, the effect of broadband does not vary in the

intensity of either routine or non-routine manual tasks. Quantile regressions measuring

26

the effect of broadband on the full wage distribution suggest that broadband increases

within-firm wage inequality.

Additionally, I am able to compare the returns of broadband adoption for new and

existing employees. I find that both new and existing employees benefit from broadband

adoption, which suggests the effect of broadband on wages is not solely the result of

recruiting better employees post-adoption. Overall employment increases 5–7 percent

following broadband adoption.

The results are useful for policymakers evaluating the potential impacts of public

investment in broadband infrastructure. Such investments are often predicated on the

hypothesis that high-speed Internet spurs economic and wage growth despite limited re-

search on this topic. I show that workers do not equally share the gains from broadband

adoption; workers engaged in higher paid occupations that require non-routine cogni-

tive tasks experience larger gains from adoption than workers in occupations intensive

in routine cognitive tasks.

27

2The Impact of Mass Shootings on Gun Policy

This chapter is co-authored with Michael Luca and Deepak Malhotra.

2.1 Introduction

Recent decades have witnessed a series of high-profile mass shootings throughout the

United States in towns ranging from Newtown, CT to Killeen, TX. While most homi-

cides receive little attention from the general public, mass shooting incidents are ex-

tremely salient. Nonetheless, a common and frequently articulated view is that despite

extensive discussion about mass shootings, they have little influence on policymaking.

Should we expect policymakers to propose new legislation in the wake of a mass shoot-

ing? Given that the vast majority of gun deaths do not result from mass shootings, it

28

would be difficult to reconcile large responses to mass shootings with basic models of op-

timal policy aimed exclusively at reducing gun violence. However, mass shootings may

have another effect—bringing attention to the issue of gun violence. Mass shootings

potentially lead to policy changes by focusing attention on gun violence, even if they

do not provide new information or change politicians’ preferences (which are generally

static and aligned with party preferences). Political scientists have noted the fact that

issues tend to rise and fall within a policy agenda, creating periods in which specific

policies change very rapidly and other periods in which they do not change at all (Baum-

gartner and Jones, 1993; Kingdon, 1984). In the context of gun violence, events like the

Columbine shooting have lead to both calls for new restrictions on guns and vehement

reaction from gun rights groups opposed to such changes (Goss, 2006; Spitzer, 2012).

More generally, mass shootings may create “policy windows” during which legislatures

become receptive to change—potentially due to shifts in the attention of constituents.

But the extent to which this occurs, and the direction of resulting changes are empirical

questions.

In this paper, we explore the impact of mass shootings on gun policy, constructing a

dataset of all U.S. gun legislation and mass shootings over a period of 25 years (1989–

2014), combining data from a variety of media and government sources. We begin by

looking at the extent of deaths resulting from mass shootings relative to other gun

deaths. Overall, there are more than 30,000 gun related fatalities in the United States

per year. Roughly 56 percent of these are suicides and 40 percent are homicides. The

29

remaining four percent are accidents or incidents of undetermined intent. Mass shoot-

ings accounted for about 0.13 percent of all gun deaths and 0.34 percent of gun murders

between 1989 and 2014.

Because mass shootings are salient and plausibly random occurrences, we are able

to implement a difference-in-differences strategy around the timing of mass shootings

to estimate their causal impact on gun regulation. Specifically, we compare gun laws

before and after mass shootings, in states where mass shootings occur relative to all

other states.

We then present three main findings about the impact of mass shootings on policy.

First, mass shootings evoke large policy responses. A single mass shooting leads to

an approximately 15 percent increase in the number of firearm bills introduced within

a state in the year after a mass shooting. This effect is largest after shootings with

the most fatalities and holds for both Republican-controlled and Democrat-controlled

legislatures.

Second, although mass shootings account for a small portion of all gun deaths, they

have an outsized influence relative to other homicides. Our estimates suggest that the

per-death impact of mass shootings on bills introduced is about 80 times as large as the

impact of gun homicides in non-mass shooting incidents.

Third, when looking at enacted laws, the impact of mass shootings depends on the

party in power. A mass shooting increases the number of enacted laws that loosen gun

restrictions by 75 percent in states with Republican-controlled legislatures. We find no

30

significant effect of mass shootings on laws enacted when there is a Democrat-controlled

legislature.

These findings contribute to the empirical literature that uses a political economy

lens to explore the determinants of policymaking (Makowsky and Stratmann, 2009;

Bardhan and Mookherjee, 2010). Our results show that salient events—such as mass

shootings—can lead to significant policy responses. Moreover, policymakers seem to

use mass shootings as an opportunity to propose bills that are consistent with their

ideology. This helps to shed light on the role of attention and salience in shaping policy

and the interaction between issue salience and existing political preferences in shaping

the degree and direction of enacted policies.

2.2 Background and Data

As described above, out of the roughly 30,000 annual gun deaths in the United States,

fewer than 100 occur in mass shootings. For the purpose of this paper, we define a “mass

shooting” as an incident in which four or more people, other than the perpetrator(s),

are unlawfully killed with a firearm in a single, continuous incident that is not related

to gangs, drugs, or other criminal activity. This definition closely matches the one

used by Krouse and Richardson (2015) and the FBI’s definition of “mass murder” as

four or more murders “occurring during the same incident, with no distinctive time

period between the murders…typically involv[ing] a single location” (Morton and Hilts,

31

2008). We further restrict our analysis to cases where at least three of the fatalities were

individuals unrelated to, and not romantically involved with, the shooter(s). We include

spree murders—homicides at multiple locations without a significant pause between

incidents—if they result in four or more deaths.

We assemble a list of mass shootings since 1989 from a variety of government and

media sources because there is no single, comprehensive government database of mass

murders. We extract all gun-related mass murders (four or more dead) that are not

felony related from the FBI supplementary homicide reports (SHR). We then verify each

incident in the SHR using media accounts; the SHR may contain errors in which separate

homicides in a month are reported as a single incident, which is why it is necessary to

verify the incidents with media coverage. Participation in the SHR program is voluntary

and many law enforcement agencies do not report detailed data to the FBI. We therefore

supplement the FBI data with mass shootings gathered from media accounts or compiled

by other researchers and journalists interested in the topic. We combine the SHR data

with mass shootings collected by the Mass Shootings in America (MSA) project at

Stanford University (Stanford Geospatial Center and Stanford Libraries, 2015) and a

list created by USA Today (2013). For each shooting, we determine the event location

as well as the number of victim fatalities and injuries. We also classify shootings based

on the relationship (if any) between the alleged shooter(s) and victims. Previous work

on mass shootings (Duwe, 2007; Krouse and Richardson, 2015) distinguishes between

public mass shootings that occur in places frequented by the public, felony-related

32

murders, and familicide. We categorize shootings by whether they are public events

or primarily related to domestic conflicts, and we focus on incidents in which at least

three people not related or romantically involved with the shooter died. This restriction

filters out family killings in residences as well as family-related murders in public places.1

Figure 2.1 shows the number of incidents and fatalities in mass shootings by year. The

data show a slight upward trend in the number of incidents and fatalities over time, but

both incidents and fatalities vary substantially from year to year.

2.2.1 Gun Legislation

State governments are the primary regulators of firearms. Federal laws establish a

minimum level of gun control, which is then augmented to varying degrees by state

and local policies. Federal government has limited commerce, the possession of guns by

potentially dangerous individuals, and some types of firearms and ammunition. States

decide a variety of gun policies ranging from who can purchase and possess a gun to

what types of guns are allowed in different situations to how guns should be stored

and what types of training should be undertaken by gun owners. Local ordinances

can also restrict firearm possession and use, but state statutes enacted in the past few

decades have limited the importance of local government in this arena by preempting

local regulations.

1A 2006 shooting at a church in Louisiana is one example. A man killed his wife and in-lawswhile abducting her and their children from a church. Only the wife’s family was present atthe church during the shooting.

33

Figure 2.1: Mass Shootings and Fatalities by Year, 1989–2014

0

20

40

60

80

Fata

litie

s

1989

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

Mass Shooting Fataliites

2468

1012

Num

ber o

f Sho

otin

gs

1989

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

Year

Mass Shooting Incidents

Note: The upper panel shows the number of fatalities in mass shootings in which at leastthree people not related or romantically connected to the shooter were killed. The bottompanel shows the number of these incidents. Washington, D.C. is not included in the sample.

34

We create a comprehensive dataset of gun legislation in all fifty states using the bill

tracking reports service from LexisNexis, which includes all bills introduced in state

legislatures since at least 1990 with a synopsis and timeline of each bill’s progress. This

allows us to determine whether bills pass the legislature and become law. We identify

firearm bills by searching for the firearm-related terms “firearm”, “handgun”, “pistol”,

“revolver”, “rifle”, “shotgun”, “long-gun”, and “assault weapon.” We identify 20,409

firearm bills and 3,199 laws between 1990 and 2014. In other words, there were 20,409

proposals introduced and 3,199 laws passed in the 25 year sample period across all fifty

states. This includes laws that loosen or tighten gun restrictions, and many that do

neither or both. We exclude resolutions, executive orders, and ballot initiatives from

the analysis.2

To explore whether gun control is tightened or loosened after mass shootings, we

hired eight people to manually code the summary of bills that became law. Coders

were given instructions explaining how to code legislation, but were otherwise blind

to the topic and design of the study. We presented bill summaries from LexisNexis

to coders in randomly chosen groups of 50. Two people coded each summary, and

no coder saw the same summary multiple times. For each summary, coders decided

whether the bill was tightening (stricter gun control), loosening (weaker gun control),

uncertain (insufficient information), both tightening and loosening, or neither tightening

2Legislators in some states first submit ideas for bills in the form of a draft request or similardocument. We exclude these from our analyses because they result in double counting somelegislation. We instead focus only on actual bills.

35

nor loosening (neutral). There were therefore five possible labels for bills: tighten,

loosen, both, neutral, or uncertain. Appendix C shows example bill summaries and

their expected labels.

To cross-validate (and incentivize) the bill coding, we coded a small fraction of bills

ourselves as a baseline comparison point. For this process, we blinded ourselves from

any information about when or where the bill was proposed. We then used our scores

to assess the quality of coders. Specifically, each group of 50 bills given to a coder

contained five bills that we had also coded (they did not know which bills were and

weren’t coded, and did not have access to any of our assessments of whether a bill was

looser or tighter). Coders were paid up to a 50 percent bonus based on the extent to

which their coding matched ours (which we simply told them was a “gold standard” of

known codes).

Across all five categories, coders agreed with each other 52 percent of the time (the

agreement rate would be 20 percent by chance) and agreed with the gold standard 71

percent of the time. Coders performed worst on the neutral category, and best on the

tighten-only and loosen-only categories; when a bill tightens gun control (according to

the gold standard), coders agree on tightening 67 percent of the time, and when a bill

loosens gun control, coders agree on loosening 60 percent of the time.

Most importantly for the purposes of our analysis: when coders agree with each other

on tightening, they also agree with our coding 93 percent of the time; when coders agree

on loosening, they are consistent with our scores 91 percent of the time. When analyzing

36

Figure 2.2: Comparison of Legislation Introduced by Political Party

l

l

l

l

ll

l

l

l

l

l

l

5

10

15

20

0.25

0.50

0.75

1.00

Bills Introduced Laws Passed Tightening Laws Loosening Laws

Mea

n Le

gisl

atio

n pe

r Ye

ar

Legislature

l

l

l

Republican

Democrat

Split

Note: Points represent the mean and lines are 95 percent confidence intervals. Legislaturecontrol means one political party includes both chambers of the legislature. The counts oftightening and loosening laws are based on laws with coder agreement (see section 2.2.1 foran explanation of coding legislation).

the direction of policy change, we leverage this high degree of reliability by restricting

our analysis to bills on which coders agreed that the law was designed to tighten or

loosen gun control. Because states can pass either, none, or both types of laws in a year,

our dependent variable is the count of laws in each direction.

Figure 2.2 shows mean bills introduced, laws enacted, and tightening and loosening

laws by political control of the state legislature. Republicans enact more laws loosening

gun control, and fewer laws tightening gun control, than Democrats. Republican, Demo-

cratic, and split legislatures enact a similar number of total gun laws. The coders who

classified the legislation were only given summaries of each bill; they were not provided

with the state, year, or any information on political affiliation.

37

2.2.2 Control Variables

While our empirical strategy allows us to control for all time invariant factors that may

affect gun legislation, we also add time varying controls. These include economic and

demographic factors such as unemployment, divorce rates, and rates of military service.

We also control for institutional differences between legislatures. First, we control for

the number of lawmakers as a measure of legislature size. Larger legislatures consider

more bills. Second, we create a dummy for legislatures that held a regular session in a

given year because not all legislatures meet annually. Third, we control for whether bills

in each year carryover into subsequent sessions; some chambers allow for carryover while

others kill all unpassed bills at the end of each session. Fourth, we control for years in

which bills were restricted to specific topics; seven states restrict the scope of legislation

(e.g. appropriations only) in specific years. Fifth, we control for the political party in

power. Table 2.1 contains summary statistics for all variables used in the analyses.

2.3 Methodology

We implement a difference-in-differences strategy that compares gun laws before and

after mass shootings, in states where mass shootings occur relative to all other states.

Our dependent variables are counts of bills or enacted laws at the state-year level. We

38

Table 2.1: Summary Statistics

Variable mean sd p5 p10 p50 p90 p95

LegislationBills introduced 16.3 22.0 0 1 10 38 53Laws enacted 2.56 3.35 0 0 1 6 9Tightening laws 0.70 1.29 0 0 0 2 3Loosening laws 0.25 0.62 0 0 0 1 1

Gun ViolenceMass shooting 0.12 0.32 0 0 0 1 1Fatalities 0.72 2.40 0 0 0 4 5Gun homicide rate 3.76 2.55 0.72 0.98 3.42 7.40 8.65

Political ControlsDemocratic legislature 0.42 0.49 0 0 0 1 1Republican legislature 0.34 0.47 0 0 0 1 1Republican governor 0.53 0.50 0 0 1 1 1

Institutional ControlsRegular session 0.94 0.24 0 1 1 1 1Bill carryover 0.27 0.44 0 0 0 1 1Limited leg. topic 0.06 0.24 0 0 0 0 1Legislature size 148 59.3 62 82.5 144 200 236

Demographic ControlsElderly (65+) % 12.9 2.0 9.8 10.7 13.1 15.2 15.7Under 25 % 35.1 2.7 31.4 32.2 34.8 38.0 39.5Black % 10.3 9.5 0.6 0.8 7.4 26.4 30.1Hispanic % 8.3 9.2 0.8 1.2 5.1 20.3 29.9Unemployment % 5.7 1.9 3.1 3.5 5.4 8.1 9.3Income per capita 19.1 3.3 14.1 15.0 18.7 23.3 25.8High school % 85.2 5.2 75.7 78.4 86.1 91.2 92.0Veteran % 11.8 2.4 7.9 8.8 11.8 15.0 16.1Divorced % 11.8 1.8 8.9 9.5 11.8 14.1 14.7

Note: Sample includes 1,250 state-year observations.See appendix B for a list of variable definitions.

39

study the effect of mass shootings using Poisson regressions with conditional mean:

E[ys,t | αs, λt, Shoots,t−1, Xs,t

]= exp

(αs + λt + β′Shoots,t−1 + γ′Xs,t

)where ys,t is a count of bills introduced or laws enacted in state s and year t; αs and

λt are state and year fixed effects; Shoots,t−1 is either an indicator for states with a

mass shooting or the fatality count in mass shootings, and Xs,t is a vector of time-

varying political, economic, and demographic factors. We estimate the parameters via

maximum likelihood by conditioning on the sum of ys,t within states and including year

indicators.

Our identification allows us to measure the impact of a mass shooting within that

state, controlling for other changes that are happening at the national level. Because

this identification strategy does not identify national responses to mass shootings (which

would be absorbed by our year effect), our estimates of changes in gun policy may under-

estimate the total impact of a mass shooting. We can account for potential spillovers

into neighboring states, but do not see significant spillover effects; these results are

presented in Appendix D.

40

2.4 Results

2.4.1 The Effect of Mass Shootings on Gun Bill Introductions

Table 2.2 shows that a mass shooting leads to a 15 percent increase in firearm bills

introduced. For the average state, this amounts to an additional 2.5 firearm bills in-

troduced in the year following a mass shooting. Mass shootings with more deaths lead

to larger effects. On average, each additional death in a mass shooting leads to a 2.5

percent increase in the number of gun bills introduced. This result holds both for

Republican-controlled and Democrat-controlled legislatures.3

2.4.2 Comparing Mass Shootings with Non-Mass Shootings

Table 2.3 shows that fatalities resulting from mass shootings lead to much larger in-

creases in gun bill introductions than gun homicides in non-mass shooting incidents.

We estimate the models in this table using mass shooting fatalities and ordinary gun

homicides per 100,000 people to facilitate comparison between the two types of murder.

It would take approximately 80 people dying in individual gun homicide incidents to

have as much impact on bills introduced as each person who dies in a mass shooting.

Our estimates imply that, on average, a single mass shooting has as much impact on

the number of bills proposed as would a 270 percent increase in the number of gun

3Results on bills proposed broken down by political affiliation are available upon request.

41

Tabl

e2.

2:Eff

ecto

fMas

sSho

otin

gson

Gun

Bill

Intro

duct

ions

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Mas

ssh

ootin

g0.

154∗

0.08

70.

155∗

∗0.

147∗

∗

(0.0

88)

(0.0

74)

(0.0

65)

(0.0

62)

Fata

litie

s0.

023∗

0.02

0∗∗

0.02

3∗∗∗

0.02

3∗∗∗

(0.0

13)

(0.0

09)

(0.0

07)

(0.0

07)

Reg

ular

sess

ion

1.94

6∗∗∗

1.93

1∗∗∗

1.95

6∗∗∗

1.93

9∗∗∗

(0.3

64)

(0.3

80)

(0.3

55)

(0.3

70)

Bill

carr

yove

r0.

500∗

∗∗0.

504∗

∗∗0.

498∗

∗∗0.

502∗

∗∗

(0.1

30)

(0.1

31)

(0.1

38)

(0.1

40)

Lim

ited

leg.

topi

c-0

.665

∗∗∗

-0.6

71∗∗

∗-0

.632

∗∗∗

-0.6

39∗∗

∗

(0.2

40)

(0.2

51)

(0.2

25)

(0.2

36)

Legi

slatu

resiz

e0.

007∗

∗0.

008

0.00

7∗∗

0.00

8(0

.004

)(0

.005

)(0

.003

)(0

.005

)D

em.l

egisl

atur

e-0

.136

∗-0

.134

∗

(0.0

74)

(0.0

76)

Rep

.leg

islat

ure

0.06

70.

071

(0.0

64)

(0.0

64)

Rep

.gov

erno

r-0

.022

-0.0

14(0

.054

)(0

.053

)

Dem

ogra

phic

Con

trol

s•

••

•St

ate

Fixe

dEff

ects

••

••

••

••

Year

Fixe

dEff

ects

••

••

••

N1,

250

1,25

01,

250

1,25

01,

250

1,25

01,

250

1,25

0

Not

e:T

hede

pend

ent

varia

ble

isth

enu

mbe

rof

firea

rm-r

elat

edbi

llsin

trod

uced

inth

est

ate

legi

slatu

re.

Rob

ust

stan

dard

erro

rscl

uste

red

byst

ate

inpa

rent

hese

s.∗p<

.10,∗

∗p<

.05,

∗∗∗p<

.01

42

Table 2.3: Comparing Mass Shootings and Non-Mass Shootings

(1) (2) (3) (4)

Mass shooting fatalities/100,000 1.678∗∗∗ 1.332∗∗∗ 1.303∗∗∗ 1.316∗∗∗(0.428) (0.240) (0.223) (0.202)

Ordinary gun homicides/100,000 -0.007 0.017 0.014 0.017(0.024) (0.032) (0.032) (0.035)

Political Controls •Institutional Controls • •Demographic Controls •State FE • • • •Year FE • • •N 1,250 1,250 1,250 1,250

Note: The dependent variable is the number of firearm-related bills introduced inthe state legislature. Robust standard errors clustered by state in parentheses. MassShooting Fatalities/100,000 is the number of deaths in mass shootings per 100,000 stateresidents. Ordinary gun homicides/100,000 is the number of gun homicides not in massshootings per 100,000 state residents.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

homicides in a state. Given the average number of gun homicides per year is roughly

260 per state, this would be equivalent to an additional 448 homicides per state-year.

2.4.3 The Role of Political Party on Laws Enacted

As mentioned previously, the two major political parties in the United States differ dra-

matically in their stances on how restrictive gun policy should be, with the Republican

Party favoring fewer gun restrictions.4 To look at the impact of political parties on gun

policy, we restrict our analysis to enacted laws, all of which were coded for whether

they loosened or tightened gun restrictions (see data description for more details).

Table 2.4 shows the effect of mass shootings interacted with Democrat and Repub-

4See https://www.gop.com/platform/ and https://www.democrats.org/party-platform.

43

lican control of state government (divided government, in which the legislature is not

controlled by a single party, is the omitted group). The results show that Democrats

and Republicans respond differently to mass shootings.

When there is a Republican-controlled legislature, mass shootings lead to more firearm

laws that loosen gun control. A mass shooting in the previous year increases the number

of enacted laws that loosen gun restrictions by 75 percent in states with Republican-

controlled legislatures. When there is a Democrat-controlled legislature, mass shootings

lead to a statistically insignificant reduction in laws that loosen gun control. We find

no significant effects of mass shootings on laws that tighten gun restrictions, but the

estimates are imprecise. Summing across all legislatures (Republican, Democrat, and

split), there is roughly a 10 percent increase in laws enacted after a mass shooting, but

this estimate is imprecise and statistically insignificant (Appendix E).

2.4.4 Robustness Checks

In this section, we present four sets of robustness checks. First, we provide support for

the exogeneity of mass shootings. Second, we show that our main results are robust to

the inclusion of state-specific time trends. Third, we perform a falsification exercise in

which we use randomly generated placebo shootings instead of actual shootings; we show

there are no effects using the placebo shootings, providing support for our identification

strategy. Fourth, we individually drop each state from the sample and re-estimate the

models to ensure our effect is not driven by a single state or shooting event.

44

Tabl

e2.

4:M

assS

hoot

ings

and

Enac

ted

Laws

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Mas

ssh

ootin

g-0

.030

0.24

6(0

.101

)(0

.181

)Fa

talit

ies

0.01

5∗0.

008

(0.0

08)

(0.0

22)

Shoo

ting×

Rep

.leg

islat

ure

-0.0

170.

733∗

∗∗

(0.2

38)

(0.2

55)

Dem

.leg

islat

ure

0.03

7-0

.250

(0.1

40)

(0.4

02)

Split

legi

slatu

re-0

.216

0.16

8(0

.262

)(0

.340

)Fa

talit

ies×

Rep

.leg

islat

ure

0.01

80.

152∗

∗∗

(0.0

50)

(0.0

31)

Dem

.leg

islat

ure

0.01

4-0

.047

(0.0

15)

(0.0

54)

Split

legi

slatu

re0.

015

-0.0

18(0

.013

)(0

.018

)D

em.l

egisl

atur

e0.

110

0.10

50.

070

0.10

5-0

.318

-0.3

03-0

.253

-0.2

59(0

.147

)(0

.148

)(0

.171

)(0

.165

)(0

.200

)(0

.198

)(0

.229

)(0

.222

)R

ep.l

egisl

atur

e0.

165

0.18

00.

134

0.17

80.

494∗

∗∗0.

496∗

∗∗0.

402∗

0.31

8(0

.137

)(0

.137

)(0

.147

)(0

.142

)(0

.187

)(0

.190

)(0

.213

)(0

.202

)R

ep.g

over

nor

-0.0

46-0

.045

-0.0

47-0

.045

-0.1

13-0

.111

-0.0

93-0

.084

(0.0

86)

(0.0

85)

(0.0

85)

(0.0

83)

(0.1

68)

(0.1

67)

(0.1

66)

(0.1

67)

Inst

itutio

nalC

ontr

ols

••

••

••

••

Dem

ogra

phic

Con

trol

s•

••

••

••

•St

ate

Fixe

dEff

ects

••

••

••

••

Year

Fixe

dEff

ects

••

••

••

••

N1,

250

1,25

01,

250

1,25

01,

175

1,17

51,

175

1,17

5

Not

e:T

hede

pend

ent

varia

ble

isth

enu

mbe

rof

firea

rm-r

elat

edla

ws

enac

ted

(i.e.

bills

that

beca

me

law

).R

obus

tst

anda

rder

rors

clus

tere

dby

stat

ein

pare

nthe

ses.

∗p<

.10,∗

∗p<

.05,

∗∗∗p<

.01

45

2.4.4.1 Determinants of Mass Shootings

Our ability to identify the causal impact of mass shootings on policy rests on the assump-

tion that they are plausibly exogenous to other factors that would drive gun control in

a given year. Given the erratic nature of mass shootings, this is a plausible assumption.

Nonetheless, one might be concerned that both mass shootings and gun policy are being

driven by a third variable. To provide support for our assumption and interpretation,

we regress an indicator for whether a mass shooting occurs on economic, demographic,

and policy variables.

Consistent with the assumption that mass shootings are exogenous with respect to

potential confounds, the results in Appendix F show that, out of 32 variables we con-

sider, only unemployment is significantly associated with a higher probability of mass

shootings. Because higher unemployment is also associated with a reduction in gun bill

introductions (Table 2.2), the potential bias of this would work in the opposite direction

of our finding—making it unlikely that this is driving our results. To further support

our interpretation, we control for unemployment in all models. Importantly, bills in-

troduced, laws enacted, and major gun policies do not predict future mass shootings

(Appendix F).

46

2.4.4.2 State-Specific Time Trends

Another potential concern is that states have differential trends in mass shootings, and

that these trends correlate with gun regulations, which would violate the parallel trends

assumption. As a robustness check, we run our main specifications with state-specific

trends. Appendix G shows the results of re-estimating the models in Tables 2.2 and 2.3

with state-specific time trends. The inclusion of state-specific trends does not change

our main results from Tables 2.2 and 2.3. We are unable to estimate models with state-

specific trends for our analyses of tightening and loosening laws because the likelihood

function is discontinuous when including the additional parameters due to some states

having very few laws that we can identify as tightening or loosening. We can, however,

conduct a placebo analysis to address any residual concerns.

2.4.4.3 Placebo Tests

We perform a falsification exercise based on the insights of Bertrand, Duflo, and Mul-

lainathan (2004) and Donald and Lang (2007).

Specifically, we randomly assign placebo mass shootings to state-years in which no ac-

tual shooting occurred with probability equal to each state’s frequency of shootings, and

randomly draw a fatality count from the empirical distribution of fatalities. Appendix H

shows percentiles of the test statistic based on 1,000 repetitions of this procedure and

our actual test statistics from Tables 2.2 and 2.4. The results suggest our tests do not

47

over-reject the null hypothesis that mass shootings have no effect on gun policy.

2.4.4.4 Excluding Individual States

To ensure our results are not driven by a single state or shooting, we separately remove

each state from the sample and re-estimate the models. Appendix I presents graphs

of the resulting 50 coefficients for the effect of mass shootings on bills and laws, and

coefficients for the Republican and Democrat interaction terms in our analysis of laws

that tightened or loosened gun policy. The results show that dropping individual states

has little effect on our estimates.

2.5 Discussion

Mass shootings account for a small fraction of gun deaths in the United States, but have

a significant impact on gun policy. More gun laws are proposed in the year following

a mass shooting, a result that holds for both Republican- and Democrat-controlled

legislatures. Notably, mass shootings have much larger effects on policy, per fatality,

than ordinary gun homicides.

However, we also find evidence that Democrat- and Republican-controlled legislatures

differ significantly when it comes to enacting gun laws. Republicans are more likely to

loosen gun laws in the year after a mass shooting. The effect for Democrats, which tends

toward less loosening of gun restrictions after a mass shooting, is statistically insignif-

48

icant. This is consistent with survey evidence suggesting that even when a majority

supports a gun control proposal, those opposed to increased gun control are more likely

to take actions like writing a letter or donating money to support their side (Schuman

and Presser, 1981).

Our results are consistent with qualitative research that has hypothesized the pos-

sibility of mass shootings precipitating change. For example, Godwin and Schroedel

(1998) argue that the Stockton schoolyard massacre in 1989 led to the enactment of

California’s assault weapons ban. We find empirical evidence that sporadic events such

as mass shootings can lead to major policy changes. This raises the question of other

factors that might drive policy, and conditions under which we might expect such ef-

fects. For example, Egan and Mullin (2012) show that extreme temperatures influence

beliefs about global warming in the short-term. Might we expect a greater impact of

random events in some policy contexts (e.g. the effect of a terrorist attack) than in

others (e.g. the effect of an Ebola outbreak)?

Our findings raise a number of additional questions, and suggest several directions for

future research. First, our estimates focus on the impact on policy within the state in

which each shooting took place. Some mass shootings get national media attention and

potentially affect policy nationwide, which would not be identified by our fixed effects

model. One direction for future research is to develop strategies to identify national

responses. With respect to our findings, this suggests that the total impact of mass

shootings on gun policy may be even larger than our estimates.

49

Second, future research might further explore the role of salience in shaping policy by

examining the conditions under which events are more or less influential. For example,

some types of events (e.g. school shootings) may have larger effects than others, some-

thing we could not test given the relative infrequency of such events. Salient events

might have a greater impact if they occur at a time when few other events are com-

peting for media attention (Eisensee and Strömberg, 2007), or during elections, when

public attention is more focused on such issues (Bouton et al., 2014). Finally, the role

of interest groups that try to promote their preferred policies in the aftermath of such

events deserves further exploration.

Third, future research might directly explore the preferences of politicians. Do Re-

publican legislatures loosen gun restrictions because Republican politicians themselves

prefer looser restrictions or due to pressure from constituents or interest groups? For

example, if constituent preferences are driving results, we might expect that results

differ in areas with high versus low rates of gun ownership. To provide exploratory

evidence, Appendix J shows the results from adding a proxy for gun ownership to the

models. Following Cook and Ludwig (2006), we calculate the percentage of suicides

that are firearm related as a proxy for gun ownership and interact this variable with

the mass shooting indicator. The coefficient on this variable is not significant either in

isolation or when added to the specification with all political interactions. This suggests

that the tendency of Republicans to loosen gun control is not entirely driven by high

rates of gun ownership (and presumably high rates of support for less gun control), but

50

represents a distinct effect of political affiliation.

Fourth, there is a large literature on the impact of gun policies on crime (Abrams,

2012; Duggan, 2001; Ludwig and Cook, 2000, 2003), which has yielded mixed results.

The relationship we find between mass shootings and gun policy raises the possibility

of using mass shootings as an instrumental variable to study the impact of gun laws

on gun deaths. Unfortunately, in our sample, mass shootings are not a sufficiently

strong instrument to estimate the effects of gun policy on gun deaths, due to their

relative infrequency. (Appendix K presents results of this analysis.) This leaves open

the possibility of using salient and plausibly random events to instrument for policy

changes in future research.

Our findings suggest that while much attention has been rightfully devoted to un-

derstanding the impact of policy, there is a lot to be learned from exploring the de-

terminants of policy change as well. We find that even random and infrequent events

that account for a relatively small portion of total societal harm in a domain might

nonetheless be crucial levers for policy consideration and change. This does not imply

that politicians and policy makers are over-reacting; it may be that on issues where

there is usually political deadlock, salient events create opportunities for change that

has been sought all along. Whether these changes reflect appropriate responses to the

problem remains an open question.

51

3Internal Labor Markets in Multi-business Firms

This chapter is co-authored with Jasmina Chauvin.

3.1 Introduction

In recent years, an active stream of research has developed around the theory of resource

redeployment, the view that firms can generate excess value by actively managing their

resources, withdrawing them from some business units and reallocating them to others

in response to changing conditions (Helfat and Eisenhardt, 2004; Levinthal and Wu,

2010; Sakhartov and Folta, 2014, 2015; Lieberman, Lee, and Folta, 2016). The theory

is attractive because it is able to explain a potential source of competitive advantage in

diversified firms as well as diversification decisions.

52

Despite these theoretical advances, with the exception of studies on internal capital

markets, we still have little evidence regarding how firms manage their internal resource

pools. Even simple descriptive statistics for the prevalence of resource redeployment

in multi-business firms are needed (Folta, Helfat, and Karim, 2016). We also lack

direct tests for key features of theory, in particular, studies showing which resources

are redeployed and how firms’ organizational features enable or constrain redeployment.

One key challenge for work in this area is the paucity of internal firm data showing how

resources are redeployed.

In this paper, we leverage a rich dataset to study how firms allocate one key resource—

their human capital—though internal labor markets. As production processes have

become more skill- and service-driven, human capital is a critical resource for many firms.

Although a rich literature exists on external labor markets (e.g. reviewed in Mawdsley

and Somaya (2016)), we know much less about how human capital is allocated and

reallocated within the firm though internal labor markets.

We develop a simple framework to predict when a firm with labor needs in a focal

business unit will staff a position by redeploying a worker internally instead of hiring a

worker in the external labor market. The resource redeployment literature has tended

to assume that some feature of a resource makes it uneconomical to transact in external

markets. Our framework allows us to be precise about when and why firms would want

to reallocate workers internally versus source externally. Specifically, we model the

possibility that internal workers are distinct from workers available in external labor

53

markets because they possess firm-specific human capital. In addition, we also model

the costs of using external labor markets, such as the costs of hiring and firing.

By incorporating these realistic features of internal and external labor markets, the

framework shows that two distinct types of motivations can drive internal redeploy-

ments. One is external labor market frictions, e.g. the costs of hiring new workers and

the costs of firing existing workers. Even if internal and external workers were other-

wise homogeneous, such frictions would create incentives for firms to redeploy workers

internally in order to avoid these transactions costs and institutional voids (Khanna

and Palepu, 2000).

A theoretically different possibility is provided by the view that workers are resources,

embodiments of knowledge (Kogut and Zander, 1992; Grant, 1996). Some of that knowl-

edge may be non-codifiable (Teece, 1981) and firm-specific (Barney, 1991). Such knowl-

edge is acquired over time and cannot be easily transferred outside the firm (Groysberg,

Lee, and Nanda, 2008). In these cases, internal and external workers are not perfect sub-

stitutes. Even in a world lacking external labor market frictions, we would still observe

redeployment motivated by the desire to allocate this firm-specific resource, embodied

in the worker, to its most productive use within the firm.

Beyond these two distinct motivations for redeployment, our framework also incorpo-

rates the idea that a firm’s organization—both its corporate strategy as reflected in the

relatedness of its activities and its geography, i.e. the location of its business units—can

enable or constrain the firm’s ability to engage in worker redeployment. We model the

54

relatedness of the firms’ activities, in particular the occupational similarity of the firm’s

different industries, as an enabler of redeployment. Greater similarity increases the

probability that the type of worker needed in a focal unit is actually available elsewhere

in the firm. We model geographic distance between the origin and destination plants as

an increase in the transfer costs involved in redeploying workers, for example relocation

expenses and incentives paid to workers to encourage them to relocate.

Guided by the framework, we study the extent and drivers of worker redeployment

leveraging a rich employer-employee matched dataset made available by the Government

of Brazil, the Relação Anual de Informações Sociais (RAIS). In it, we can observe

all workers employed at a firm, as well as their movements among the firm’s plants.

For the current analysis, we select a ten percent random sample of all multi-business

firms operating in Brazil from 2004–2014. During this time period, multi-business firms

accounted for 14–17 percent of the total formal sector labor force of Brazil.

In stylized facts, we observe that Brazilian multi-business firms source a substantial

share of their labor needs internally. On average 12.1 percent of workers hired in any

year come from other establishments of the same firm. At any point in time, redeployed

workers represent 5.5 percent of a plant’s workers. Among workers leaving an establish-

ment, 11.8 percent move to jobs within the same firm. This percentage is even higher

when firms close an establishment; 21.8 percent of workers in establishments that are

closing down move to new positions within the same firm.

We next analyze worker-level models to gain insight into which employees are more

55

likely to be redeployed and thus infer the motivations of redeployment. Two findings

emerge consistently. First, comparing otherwise similar workers employed at the same

plant and occupation group in a year, workers with more firm-specific experience are

more likely to be redeployed. All else equal, a worker with the average level of firm

experience (2.9 years) has a 7.4 percent higher likelihood of being redeployed compared

to a worker with no firm-specific experience. Second, within a given plant, workers

higher in the organizational hierarchy are more likely to be redeployed. In particular,

on average, 8.2 percent of an establishment’s managers are redeployed in any year. This

is nearly double the 4.4 percent of service and production workers that are redeployed. If,

as commonly thought, valuable firm-specific human capital tends to reside with workers

higher up in the hierarchy, these results are consistent with the hypothesis that firms

use redeployment as a tool to reallocate valuable human capital resources.

In order to better tease apart a hypothesis of firm-specific human capital from other

potential drivers of worker redeployment—for example, external labor market frictions

(e.g. higher search and information frictions for managers) or redeployments due to work-

ers’ personal motivations—we test whether internally redeployed workers earn a wage

premium over otherwise similar workers hired in the external labor market in the same

plant, occupation, and year. We find strong evidence of wage premia to internally rede-

ployed workers. Specifically, internally redeployed workers enjoy a nine percent higher

contractual wage compared to otherwise similar workers hired into the same occupation

and establishment though the external labor marker. Moreover, this premium is small

56

for internal workers without firm-specific experience (i.e. workers hired and immediately

redeployed) and rising steeply in a worker’s years of firm-specific experience. These find-

ings also are consistent with the existence of productivity-enhancing firm-specific human

capital which allows internal workers to generate (and capture) excess returns. They

are less consistent with pure frictions in external labor markets or moves motivated by

workers’ personal preferences.

Exploring to what extent large firing costs may be driving redeployments, we find

that while plant exits are associated with more workers being redeployed, overall, only 11

percent of the redeployments that we observe are concurrent with plants shutting down.

We do not find that plants shutting down become more likely to redeploy “blue-collar”

workers who may otherwise present additional firing costs (e.g. due to unionization)

(Cestone et al., 2017). Rather we find that when a plant exits, workers highest in the

hierarchy and those with more firm-specific experience are more likely to be redeployed.

Finally, we explore how the firm’s organizational features, in particular the related-

ness of activities and its geographic footprint, affect the extent of worker redeployment.

Here we model the volume of redeployments between all possible sets of origins and des-

tinations (dyads) in a firm as a function of their industry relatedness, their geographic

distance, and proxies for differences in the growth patterns of their respective industries.

We find evidence consistent with the view that greater labor relatedness between indus-

tries facilitates the redeployment of workers while greater geographic distance between

plants strongly discourages worker redeployment.

57

Taken together, the findings point to the conclusion that internal labor markets within

multi-business firms serve as a conduit though which firm-specific human capital is

transferred among the firm’s units. We find particularly strong evidence that firms

redeploy their managerial human capital, and especially those workers with higher levels

of firm-specific experience. We find strong evidence that workers with more firm-specific

human capital earn excess rents in the form of higher wages.

The view of internal labor markets supported by our findings is quite distinct from

other prevailing views. Until recently, much of the literature of internal labor markets

focused on “vertical” labor markets, or “career ladders”—i.e. the processes through

which workers move up the hierarchy within a given firm and the ways that firms can

design appropriate promotion mechanisms for workers over their careers (e.g. Doeringer

and Piore, 1971). Recently, a literature has begun to emerge which, rather than vertical

considers the unique aspects of horizontal internal labor markets, i.e. worker moves in

multi-plant and multi-business firms. Thus far, existing studies have focused primarily

on the potential of internal labor markets to avoid frictions and rigidities of external

labor markets (Belenzon and Tsolmon, 2016) and to enable firms to adjust to unexpected

shocks, e.g. by reallocating workers from plants that are shutting down to other parts

of the firm (Tate and Yang, 2015).

In this paper we propose that, beyond these possibilities, internal labor markets can

also play the role of allocating valuable, firm-specific human capital to the parts of the

firm where it is most needed. Our view is consistent with redeployment motivated by

58

the existence of rare and valuable resources which are otherwise not available or easy to

transact in external markets. Beyond offering evidence for this alternative motivation

for internal labor market activity, our study is also unique in exploring the organi-

zational enablers of worker redeployment. Our findings suggest that the relatedness

between the different industries of the multi-business firm and the geographic proximity

of the firm’s units facilitate workers redeployment. This implies that firms for which

worker redeployment is an important part of the strategy and a source of competitive

advantage, face a trade-off between the objectives of expanding their geographic and

product boundaries and facilitating the flows of workers though the firm’s internal labor

market.

3.2 Theory and Hypotheses

The resource-based view sees the firm as a collection of “those (tangible and intangible)

assets which are tied semipermanently to the firm” (Wernerfelt, 1984, p. 172). An

important emphasis in this view is embodied in the word tied, which implies that these

assets have features that make them difficult to transact in the open market. If assets

are homogeneous or perfectly mobile, they are not a resource, which are those assets

that are valuable, rare, imperfectly imitable, and not substitutable (Barney, 1991).1

1The resource-based view is one of serval theories of the firm. In alternative theories—forexample, firms as a “nexus of contracts” (Fama, 1980)—resource flows among the firm’s unitsare not a necessary feature.

59

As a firm learns and grows, some resources get freed up and “slack” is created (Penrose,

1959). Because slack resources are difficult to transact in the open market (Teece,

1982), assuming that they are fungible across activities and that the firm cannot expand

infinitely in its primary product, this provides an incentive the firm to diversify and thus

gives rise to the multi-business firm. Once diversified, the ability to generate synergies

through the simultaneous use of resources across multiple activities provides economies

of scope and is a source of competitive advantage for multi-business firms.

More recent additions to this literature highlight that while some internal resources

have the quality of being public goods within the firm (scale free) others are rival

and their use in one part of the firm limits their use in other parts (non-scale free)

(Levinthal and Wu, 2010). For example, a firm’s brand is a scale-free resource and can

be simultaneously leveraged in different units of the firm. However, the time and skill

of the firm’s managers is a non-scale-free resource. Although scale-free resources lend

themselves to the simultaneous use across business units and the generation of synergies,

non-scale-free resources do not.

However, firms can achieve competitive advantages in resource use even for resources

that are non-scale-free through resource redeployment (Helfat and Eisenhardt, 2004).

Rather than using resources simultaneously in two different business units, firms can

redeploy resources from activities where there are less valuable and toward activities

where they are more valuable. The benefits generated by a strategy of redeployment

are termed inter-temporal economies of scope, as they generate competitive advantage

60

through the ability to optimally adjust resource use across activities over time. Resource

redeployment can help firms exit businesses with declining prospects while lowering the

costs of starting or expanding operations in more promising areas (Lieberman, Lee, and

Folta, 2016). Note that the original business unit does not necessarily close as part of

a strategy of redeployment (Folta, Helfat, and Karim, 2016).

Although theoretically attractive, synergies and resource redeployment have been

very difficult to study empirically. A key reason is the rarity of data internal to firms

that show how they allocate resources among the different business units.2 Notable are

approaches based on the observations or resource reconfiguration within particular firms

(Karim and Mitchell, 2004), though those pose the question how generalizable strategies

are across firms. Other approaches induce redeployment by observing business unit

entries and exits (Lieberman, Lee, and Folta, 2016), though the actual flows of resources

are not observed.

In this paper, we directly observe the movement and reallocation of one important

type of resource across a large set of firms—workers. Human capital resources are one

of the three resource categories identified by Barney (1991) and include “the training,

experience, judgment, intelligence, relationships, and insight of individual managers

and workers in a firm” (Barney, 1991, original emphasis). However, not all workers

constitute resources. To the extent that a worker’s attributes are homogeneous, and

2An important exception is the literature on internal capital markets, which has documented theextensive use of internal allocation mechanisms and the relative advantages of internal versusexternal capital markets, e.g. Lamont (1997); Stein (1997); Shin and Stulz (1998).

61

thus easily substitutable, then this worker would not be considered a resource. On the

other hand, if a worker has some rare skills or has made certain firm-specific investments

and possesses firm-specific knowledge (Morris et al., 2017), then the worker constitutes

a resource. If workers possess firm-specific knowledge, they are not fully substitutable

though workers available in external labor markets. With the exception of a few types of

workers (e.g. the CEO), workers are a non-scale-free resource—their use in one activity

prevents their use in another.

How do multi-business firms decide how to optimally allocate this key resource, work-

ers, across their different business units? Both the theoretical and empirical literature

on this specific question is scarce. Existing studies have tended to focus on the internal-

to-external transitions of workers (e.g. see literature reviewed in Mawdsley and Somaya

(2016)) or internal labor markets as a means of vertically transitioning workers though

a firm’s hierarchy, i.e. “career ladders” (Doeringer and Piore, 1971).

We develop a simple model of the decision to fill labor needs in a business unit by

redeploying workers internally instead of hiring them in the external labor market. The

model incorporates the assumption that (at least some) workers have firm-specific hu-

man capital, i.e. indeed constitute a resource. The model also incorporates key features

of the theory of resource redeployment (Helfat and Eisenhardt, 2004; Sakhartov and

Folta, 2014, 2015), such as adjustment costs (in particular, industry relatedness) and

industry-level inducements, tailoring them to the specific context of internal labor mar-

kets. Finally, we incorporate the existence of external market frictions, in particular cost

62

and rigidities associated with hiring and firing of workers (Lafontaine and Sivadasan,

2009; Belenzon and Tsolmon, 2016). By explicitly modeling the choice of internal re-

deployment alongside the alternative of external market resource acquisition, we are

precise about the conditions under which internal market transactions are preferable,

which the literature has tended to not specifically address.

3.2.1 A Simple Model of Worker Redeployment

In this section, we propose a simple model to gain insights under what conditions a multi-

business firm will staff a labor need in a focal business unit via internal redeployment

versus the external labor market.

The firm’s objective in any period is to maximize the sum of profits across its business

units. We assume, for simplicity, that the firm operates two business units, one in

industry j and the other in k, such that πf = πj + πk. We assume that demand in each

business unit is exogenously given by dj = D̄j/Nj where D̄j is industry demand and

Nj is the number of firms in industry j, and that prices are perfectly competitive with

p = 1. Each business unit j requires one worker of a particular type o (think of type

as an occupation, e.g. a welder) who can produce any quantity of output at a constant

marginal cost. Labor is the only input into production and the constant marginal cost of

a unit of output is MCj = w/L̃ij where w is the wage and L̃ij is the labor productivity

of worker i in business unit j. A business unit employing worker i, thus has variable

63

profits of: πj = dj(1− wL̃ij

).3

The labor productivity of a worker of type o in business unit j is a function of two

terms: 1) the worker’s general skills (e.g. general expertise, education, experience) and 2)

the worker’s firm-specific human capital (e.g. tacit knowledge, internal social networks),

which we model as: L̃ij = hisif . Letting I represent an internal and X an external hire,

we assume that sI > 1 and sX = 1, meaning only internal workers have productivity-

enhancing firm-specific human capital. Thus, we have: L̃Iij = hisif and L̃X

ij = hi with

sif ≥ 1. In what follows, for brevity, we denote πIj as the profits generated in a business

unit when it employs an internal hire and πXj the profits generated by employing an

external hire.

A firm looking to hire a unit of labor in business unit j faces two options to fill the

position: hiring externally or redeploying the worker internally from k. Hiring for j in

the external labor market incurs a one-time hiring cost HCj (e.g. the costs of time,

search) while redeployment from k to j incurs a transfer cost TCkj . Brazilian labor

law, for example, guarantees certain rights to employees in the case of internal company

transfer to a different address, among them the payment of relocation expenses. If the

3Because we are focused on within-firm, rather than between-firm dynamics, in order to keepthe framework simple, we assume perfect competition on the demand side with cost differenceson the supply side, without specifying a general equilibrium model of competition betweenfirms where firms price at marginal cost. However, the model conclusions do not hinge onthese simplifying assumptions. The key aspects that are required for our analysis are thatfirms’ demand is determined to some extent by external industry conditions and that firms haveheterogeneous marginal costs with lower marginal costs mapping into higher profits. Specifyingconstant electricity of substitution demand with monopolistic competition among firms wouldalso deliver the conditions needed for the analysis.

64

firm has no slack in k (an assumption to be relaxed later), redeploying the worker also

implies replacing her with an external hire in the origin unit k, i.e. incurring a hiring

cost in k, HCk.

Finally, whether a worker of type o is available within the firm is a function of the

labor similarity of industries j and k, which we denote by γjk. For example, if j requires

a welder, than γjk denotes the likelihood that a welder is indeed available in business unit

k (e.g. if k is the firm’s marketing unit, this probability will be low). This probability

will range from zero to one, i.e. 0 ≤ γjk ≤ 1.4 Conditional on a worker of type o

being available within the firm (a probability that’s increasing in γij), then the firm will

choose the optimal hiring institution, H∗. The choice between hiring internally (HI)

and hiring externally (HX) is:

H∗ =

HI if πI

j + πXk − TCkj −HCk ≥ πX

j + πIk −HCj

HX otherwise

(3.1)

This inequality can also be expressed as, hire internally if:

(πIj − πX

j ) + (πXk − πI

k) ≥ TCkj − (HCj −HCk) (3.2)

Equation 3.2 shows that an internal redeployment will take place if the incremental

benefits that an internal worker generates over an external worker in destination unit j

(term 1) net of the benefits foregone by replacing an internal worker with external one

4Note that we assume that external labor markets are thick enough that a firm can always findthe worker of the needed type.

65

in the origin unit k (term 2) exceed the transfer cost net of any difference in the cost of

external hiring in j relative to k.

In the absence of prohibitive hiring costs in j, for a redeployment incentive to exist it

has to be the case that the first term of the equation is positive, πIj − πX

j > 0: internal

workers have an advantage over external ones in j. Omitting the worker subscript i, this

implies: dj(1− whsf

) > dj(1− wh ). Assuming for now that internal and external workers

with the same general skills are paid the same wage, wI = wX , this is true when sf > 1,

firm-specific human capital advantages exist. Therefore, in the absence of slack and

with zero differences in hiring frictions across locations, workers possessing firm-specific

human capital is a necessary condition for redeployment to take place.

At the same time, note that an internal worker with high levels of firm-specific human

capital will also be more valuable in the origin.5 Indeed, in order for the transfer to be

profitable, it has to be the case that the gains of the internal worker in j exceed the

opportunity cost of the internal worker at origin k. This condition is met when:

dj(1−w

hsf)− dj(1−

w

h) ≥ dk(1−

w

hsf)− dk(1−

w

h) (3.3)

which is true when dj ≥ dk—i.e. demand conditions in the destination are weakly better

than in the origin.

Thus a positive difference in demand conditions between the destination and origin

5Note that we assume that an internal worker is always at least as profitable as an externalworker in the origin business unit, i.e. sif ≥ 1 in the origin.

66

is a second necessary condition for worker redeployment to occur. These observations

allow us to formalize our first set of hypotheses:

Hypothesis 1: All else equal, a worker will more likely be redeployed the higher

their level of firm-specific human capital. (firm-specific human capital hypothesis)

Hypothesis 2: Redeployments will be higher, the more positive the industry con-

ditions of the destination relative to the origin. (inducement hypothesis)

Hypothesis 3: Redeployments will be higher, the greater the labor relatedness of

two industries. (relatedness hypothesis)

Revisiting equation 3.2, even if firm-specific knowledge and positive industry differ-

entials exist, the relative benefits of the internal worker in j have to be sufficiently large

to compensate for transfer costs. In the case of worker redeployment, transfer costs

will include things like reimbursements that a firm has to pay a worker for the cost

of moving, as well as any incremental incentives (e.g. bonuses) that the firm will pay

to encourage the worker to relocate. In general, geographic distance between plants is

likely to imply greater transfer costs, and thus all else equal, fewer redeployments. Fi-

nally, per equation 3.2, a differential in the external market hiring costs can also create

an incentive for redeployment. In particular, destinations where the external market

hiring costs are high, are likely to see more redeployments from within the firm. Seen in

a different light, the incremental advantages of an internal workers at the destination

67

can be lower for destinations where the external labor market hiring costs are high. We

formalize these observations:

Hypothesis 4: Redeployments will be will be higher the lower the geographic

distance between the origin and destination business units. (distance hypothesis)

Hypothesis 5: Redeployments will be higher the more unfavorable local labor

market conditions of the destination relative to the origin business unit. (external

labor market frictions)

We next consider a situation where the firm has slack, defined as a need of firing a

worker in the origin business unit.6 This could be due to learning-by-doing, or because

the firm is shutting down the origin business unit, for example, due to unfavorable

industry conditions.7 When a firm has a hiring need in j and a slack worker in k, it will

redeploy the worker internally if:

πIj − TCkj ≥ πX

j −HCj − FCk (3.4)

which can be rewritten as:

(πIj − πX

j ) ≥ TCkj − (HCj + FCk) (3.5)

6We assume that workers are indivisible and use the term “slack” to denote that the workeris superfluous, rather than that she has some excess capacity — i.e. we do not model thepossibility of synergies whereby a worker is active in more than one establishment at the sametime.

7While, to keep the framework simple, we have not modeled any fixed costs of operating, onecan imagine the existence of fixed overhead costs in each period, which can lead a firm to decideto shut down when demand is low.

68

Comparing equations 3.2 and 3.5, we see that the threshold for redeployment is always

lower when the firm has slack. Note also from equation 3.5, that in some scenarios, the

sum of hiring and firing costs may exceed the transfer cost and redeployment may occur

even when the right hand side of equation 3.5 is negative—i.e. when internal workers

do not have a productivity advantage over external workers in the destination unit.

Therefore, redeployments motivated by plant closures may lead to some “inefficient

hires” from the perspective of the receiving business unit, which would have been able

to attract a higher quality human capital in the absence of firing costs in the origin.

Note also that, keeping constant the left side of equation 3.5, a situation involving slack

and positive firing costs implies potential for higher levels of transfer costs, compared

to a situation of no slack. These observations lead to:

Hypothesis 6a: All else equal, a worker will be more likely to be redeployed if

their business unit is exiting. (slack hypothesis)

Hypothesis 6b: Redeployments occurring when a business unit is exiting will occur

at higher geographic distances, on average, than redeployments occurring when a

business unit is not exiting. (slack-distance trade-off hypothesis)

Finally, note that thus far, we have assumed that any productivity advantages that

internal workers generate due to their firm-specific human capital accrue to the firm via

higher profits. In reality, is it likely that firms and workers share the rents generated

via some form of Nash bargaining. Therefore, due to their firm-specific human capital,

69

internal hires may earn higher wages than external hires with the same level of general

skills: wI > wX | hI = hX .8 Combining the possibility of excess returns to firm-specific

capital with the insights reflected in the prior hypotheses we propose:

Hypothesis 7 : A redeployed worker is likely to earn higher wages compared to a

worker in the same position and comparable general skills hired externally. The

wage premium to the internal worker will be increasing in the worker’s level of

firm-specific human capital. (wage advantage hypothesis)

Overall, this simple model provides several insights into a dual role of internal labor

markets. We see that there are two distinct types on inducements to redeployment ac-

tivity: one, the desire to transfer the “best” workers to their most productive uses, for

example in response in differences industry prospects and two, the desire to reduce ad-

justment costs given slack in an existing business units. These two types of inducements

have different implications for which workers are transferred, the wage earnings of the

transferred workers, and the productivity advantage of internal versus external hires for

the firm. Although redeployments in the absence of slack incentivize the transfer of the

workers with the highest levels of firm-specific human capital, redeployments involving

slack will be associated with relatively lower levels of human capital and may even result

in some “inefficient redeployments” from the perspective of the destination unit. Thus,

8Although it’s also possible that internal hires may be willing to accept a lower wage thanexternal hires, which would also allow for instances of redeployment driven by this internalwage gap, such cases should not be part of an equilibrium as worker could always quit andreceive the (higher) market wage for their skills.

70

we see that internal labor markets can play both the function of allocating firm-specific

knowledge to its most valuable uses as well as providing an adjustment mechanisms to

weather negative shocks, with better-performing different business units absorbing slack

generated in business units that under-perform or experience a negative external shock.

3.3 Data and Empirical Strategy

Our primary data source is the Relação Anual de Informações Sociais (RAIS), a manda-

tory, annual census of all formal-sector employers and their employees in Brazil. These

data are collected by the Ministry of Labor to support various social insurance programs

and contain detailed information on the wages, occupation, and demographics of work-

ers along with the industry and location of employers. Importantly for our purposes,

RAIS is an establishment-level census with unique identifiers for each worker, establish-

ment, and firm. We can therefore link workers to firms and observe them moving both

between and within firms over time.

We take a ten percent random sample from the population of firms in RAIS that

operated establishments in multiple industries between 2004 and 2014. This results in

an initial sample of 31,428 establishments in 8,535 unique firms. The average (median)

multi-business firm has 3.6 (2) establishments in 2.1 (2) industries.

To identify instances of worker redeployment, we start with observations from the first

and last month that each worker was employed during the year and take the worker’s

71

highest paying job within each establishment-month pair. We code redeployment as

a worker switching establishments either between the first and last month of employ-

ment within a year or from one year to the following calendar year. When analyzing

redeployments, we exclude the first and last years of our sample because we cannot ob-

serve redeployments occurring between years for the initial and final sample year. This

procedure identifies 573,259 worker redeployments for 455,514 unique workers in 7,605

firms over the nine-year period from 2005 to 2013. The final sample of redeployments

contains fewer firms than the initial random sample due to the exclusion of the first and

last sample years.

Following recent literature (Sakhartov and Folta, 2014), our main measure of industry

resource relatedness is built from the similarity of industries’ occupational requirements.

Using data from the year 2000 from RAIS—five years before our sample period—we cal-

culate each of 2,331 different occupations’ share of total employment for each industry

and then calculate labor relatedness between any two industries as one minus the Eu-

clidean distance of their occupation shares. We then normalize this variable across all

the industry pairs to have mean zero and variance equal to one.

To measure the industry opportunities that may act as an inducement for redeploying

workers to activities with high returns, we use the two-year growth rate in total industry

employment. This measure assumes that greater employment growth within an industry

over time is indicative of better prospects for firms. This variable is calculated from

RAIS as the total number of unique workers with jobs within a given industry and year

72

across all firms in Brazil.

Tables 3.1, 3.2, and 3.3 show summary statistics for workers, establishments, and

destination-origin establishment pairs respectively.

We conduct two types of analyses, one at the level of individual workers and the other

at the level of business unit dyads. The first set of worker-level models take the form:

Redepikt = β + βssit +∑l

βlhit + βzzkt + ηk + τt + ϵikt (3.6)

where the dependent variable takes the value one if a worker was redeployed in the year

and zero otherwise. sit is a proxy for a worker’s level of firm-specific human capital in

year t, hit control for the worker’s general level of human capital (education, age, age

squared, gender), zkt is establishment size, η and τ are establishment and year fixed

effects, and ϵikt is a randomly-distributed error term. We perform the analysis using a

linear probability model.

The focus of the model is the coefficient on the firm-specific capital proxy, βs, which

conditional of the worker’s general skills, estimates the effect of a worker’s level of firm-

specific human capital on their probability of redeployment. Note that, due to the rich

set of fixed effects, the comparison is among workers with different levels of firm ex-

perience working in the same establishment in the same year. We employ two proxies

of a worker’s level of firm-specific human capital. The first is the worker’s years of

work experience with the firm. Our second proxy is the worker’s position in the occu-

pational hierarchy, i.e. whether the worker’s occupation falls in the director/manager,

73

Table 3.1: Worker Summary Statistics

Variable mean sd p5 p25 p50 p75 p95

Redeployment 0.04 0.18 0 0 0 0 0Log wage 0.77 0.72 0.01 0.28 0.60 1.09 2.23Firm experience 2.85 4.16 0 0 1 4 11Age 31.48 9.77 19 24 29 37 51Female 0.37 0.48 0 0 0 1 1

Occupation groupsManagers 0.03 0.18 0 0 0 0 0Professionals 0.04 0.20 0 0 0 0 0Technicians & Admin 0.27 0.45 0 0 0 1 1Service & Production 0.65 0.48 0 0 1 1 1

Education groupsBelow high school 0.40 0.49 0 0 0 1 1High school 0.51 0.50 0 0 1 1 1Higher education 0.09 0.29 0 0 0 0 1

Note: Redeployment is a dummy variable for worker redeployment in a given year.Firm experience and age are measured in years.

Table 3.2: Establishment Summary Statistics

Variable mean sd p5 p25 p50 p75 p95

Employees 66.4 323.1 1 4 11 33 249New hires 26.7 151.1 0 1 4 13 96Separations 23.0 120.9 0 1 3 12 85Closing year 0.06 0.3 0 0 0 0 1

Note: Separations refers to workers leaving the establishment. Closing year refers tothe last year an establishment operates with employees in RAIS.

Table 3.3: Destination-Origin Dyad Summary Statistics

Variable mean sd p5 p25 p50 p75 p95

Industry similarity 1.82 1.20 -0.57 1.41 2.44 2.44 2.44Distance (km) 824 816 18 203 545 1,145 2,683Difference in growth 0.00 0.12 -0.11 0.00 0.00 0.00 0.11Closing origin 0.06 0.25 0 0 0 0 1

Note: Difference in growth is calculated as the two-year employment growth rate inthe industry of the destination establishment less the same growth rate in the industryof the origin establishment. Closing origin refers to the last year that an origin operateswith employees in RAIS.

74

professional, technical and administrative personnel, or production and service worker

category. Our prior is that workers higher in the occupational hierarchy are likely to

posses rarer and more valuable firm-specific human capital.

Based on H1, the specific human capital hypothesis, we expect βs to be positive—

i.e. workers with higher levels of firm-specific human capital will have a higher likelihood

of being redeployed. We also use a slightly modified version of this model to test H6a

(the slack hypothesis) by adding to the model specified in equation 3.6 an indicator

variable for whether the worker’s current establishment exits in that year. Based on the

hypothesis, we expect the coefficient on the exit indicator to be positive and significant.

Our second worker-level model takes workers’ contractual wage as the dependent

variable. Our tests regarding the wage advantages of redeployed workers over workers

hired in the external labor market take the following form:

lnwiojt = β + βiRedepit + βssit +∑l

βlhit + θojt + ϵiojt (3.7)

The sample of observations for this model are all new employees joining business unit j

at time t, which are sourced from either the internal or the external labor market. redepit

is an indicator variable taking the value one if the worker was redeployed internally and

zero if hired externally. By including a fixed effect for each occupation-establishment-

year combination and the full set of worker controls, we are effectively comparing the

wage of an internal and an external hire entering the same occupation, in the same

establishment, in the same year with the same observable characteristics. Per H7 (the

75

wage advantage hypothesis), we expect βi to be positive. To further test whether it is

the worker’s firm-specific human capital that is driving the wage premium, we introduce

βssit. The hypothesis is that the wage premium of workers with little firm experience

will be small while βs will be positive.

The final set of models are estimated at the level of business unit dyads. We construct

the set of all possible origin- and destination- business units (dyads) within each firm

and measure the total amount of redeployments between them in each year (note that

each dyad is directional, thus a → b is not equal to b → a). We estimate the following

model:

Redepkjft = α+ β1△djkt + β2γjk + β3Geojkt + β4zjt + β5zkt + ξf + τt + ϵkjft (3.8)

where Redepkjft is the number of workers redeployed from origin business unit k to

destination j within firm f in year t, △djkt is a measure of the difference in prospects

of j and k’s industries, γjk is the industry relatedness, Geojkt is geographic distance

between the plants and the zs are establishment controls. The model also includes firm-

and year fixed effects. To test H2 (the inducement hypothesis), we focus on the sign of β1

which is expected to be positive. We test H3 (the relatedness hypotheses) by evaluating

the sign on β2 which is expected to be positive. H4 (the distance hypothesis) predicts

that the sign of β3 is negative. In other versions of this model, we also introduce an

indicator for whether the origin business unit exits at t and its interaction with the main

effects, to test H6b, whether redeployments occur at larger distances when the origin

76

Table 3.4: Percentage of Workers Redeployed by Establishment

Condition mean sd p50 p75 p90 p95 p99

Incoming redeployments as percentage ofEmployees 5.5 15.2 0.0 2.5 15.4 33.3 91.3New Hires 12.1 24.0 0.0 12.5 50.0 71.4 100New Plants 22.5 32.8 0.0 40.0 81.8 100 100

Outgoing redeployments as percentage ofEmployees 4.8 13.8 0.0 2.2 13.6 25.0 83.3Separations 11.8 23.9 0.0 12.5 44.4 68.8 100Closing Plants 21.8 35.2 0.0 33.3 100 100 100

Note: Numbers represent redeployments as a percentage of employees in each cat-egory. Incoming redeployments are workers joining an establishment from anotherestablishment owned by the same firm; outgoing redeployments are workers leavingan establishment for another establishment of the same firm. Numbers for hires andseparations (i.e. workers leaving the establishment) are conditional on anyone beinghired or exiting.

plant is exiting.

3.4 Results

Worker redeployment is pervasive. Table 3.4 shows that 12.1 percent of new hires

come from other establishments of the firm. Among workers leaving an establishment,

11.8 percent move to jobs within the same firm. This percentage is even higher when

firms close an establishment; 21.8 percent of workers in establishments that are ceasing

operations move to new positions within the same firm.

Employees higher in the organizational hierarchy and employees in professional roles

are more likely than others to be redeployed between establishments of the same firm.

Table 3.5 shows that on average, 8.2 percent of an establishment’s managers are rede-

77

Table 3.5: Percentage of Workers Redeployed by Occupation

Occupation All Years Closing Plants New Plants

Managers 8.2 27.7 41.6Professionals 6.4 30.2 34.1Technicians & Admin 5.0 23.4 25.4Service & Production 4.4 21.3 20.5Total 5.6 23.7 27.0

Note: Numbers represent redeployments as a percentage of employees in each cat-egory. Closing Plants refers to establishments in their final year of operation andnumbers represent the percentage of employees who move to another establishment ofthe same firm. New Plants refers to the first year of a new establishment and numbersrepresent the percentage of employees hired from other establishments of the samefirm.

ployed in any year. This is nearly double the 4.4 percent of service and production

workers that are redeployed. The gap in redeployment between managers and others

narrows when establishments close. On average, closing establishments redeploy 28 per-

cent of their managers, 30 percent of their professionals, and roughly 22 percent of other

workers. Managers’ increased likelihood of redeployment is suggestive evidence for the

hypothesis that firms use redeployment as a tool for reallocating valuable human capital

resources.

Consistent with the hypothesis that redeployment is increasing in firm-specific human

capital, Table 3.6 shows an additional year of experience working in a firm increases the

probability of redeployment by roughly 0.1 percentage points. About 3.5 percent of

all workers are redeployed in any given year, so a 0.1 percentage point increase repre-

sents a 2.9 percent increase in the probability of redeployment. Model 2 of Table 3.6

provides further support for the hypothesis. Workers who are more likely to have valu-

78

Table 3.6: Redeployment of Workers

(1) (2) (3) (4) (5)

Firm experience 0.0009∗∗∗ 0.0007∗∗∗ 0.0008∗∗∗(0.0003) (0.0002) (0.0003)

Managers 0.0481∗∗∗ 0.0466∗∗∗ 0.0466∗∗∗(0.0118) (0.0113) (0.0121)

Professionals 0.0091∗∗ 0.0086∗ 0.0076∗(0.0045) (0.0044) (0.0046)

Technicians & Admin 0.0069∗∗∗ 0.0064∗∗∗ 0.0059∗∗∗(0.0012) (0.0012) (0.0012)

Closing 0.1353∗∗∗ 0.1315∗∗∗(0.0321) (0.0309)

Closing ×Firm experience 0.0045∗∗

(0.0020)Managers 0.0815∗∗∗

(0.0248)Professionals 0.1137∗∗∗

(0.0313)Technicians & Admin 0.0609∗∗∗

(0.0226)High school 0.0047∗∗∗ 0.0029∗∗∗ 0.0031∗∗∗ 0.0047∗∗∗ 0.0028∗∗∗

(0.0008) (0.0008) (0.0008) (0.0008) (0.0008)Higher Ed 0.0161∗∗∗ 0.0071 0.0075∗ 0.0163∗∗∗ 0.0072

(0.0042) (0.0045) (0.0043) (0.0042) (0.0045)Age 0.0013∗∗∗ 0.0014∗∗∗ 0.0012∗∗∗ 0.0013∗∗∗ 0.0013∗∗∗

(0.0003) (0.0003) (0.0003) (0.0003) (0.0003)Age squared -0.0000∗∗∗ -0.0000∗∗∗ -0.0000∗∗∗ -0.0000∗∗∗ -0.0000∗∗∗

(0.0000) (0.0000) (0.0000) (0.0000) (0.0000)Female -0.0031∗∗∗ -0.0031∗∗∗ -0.0029∗∗∗ -0.0032∗∗∗ -0.0031∗∗∗

(0.0011) (0.0010) (0.0009) (0.0011) (0.0009)Log employment 0.0141∗∗∗ 0.0139∗∗∗ 0.0144∗∗∗ 0.0195∗∗∗ 0.0193∗∗∗

(0.0049) (0.0049) (0.0049) (0.0051) (0.0051)

Firms 7,474 7,474 7,474 7,474 7,474Observations 6,179,313 6,179,313 6,179,313 6,179,313 6,179,313

Note: Standard errors in parentheses are clustered by firm. All models include establishment andyear fixed effects. The excluded category for the education dummies is “Less than high school”; theexcluded category for the occupation categories is “Service & Production” workers.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

79

able firm-specific human capital—managers and professionals—are much more likely to

be redeployed. Controlling for worker characteristics and establishment and year fixed

effects, managers are redeployed at a rate that is nearly five percentage points higher

than the probability of redeployment for service and production workers.

Closing an establishment dramatically increases the probability that workers will be

redeployed. As hypothesized, Table 3.6 shows that closing an establishment is associated

with a 13.5 percentage point increase in the probability of redeployment; this is a more

than 300 percent increase in the probability of redeployment. Models 4 and 5 further

show that the impact of closing an establishment is even greater for employees with

more firm-specific work experience and those higher in the organizational hierarchy—

e.g. managers. Intuitively, an establishment closing represents an opportunity for the

loss of rents from firm-specific human capital. In these cases, redeployment allows the

firm to keep a worker within its boundaries and therefore maintain any benefits of

firm-specific human capital.

Workers who are redeployed (i.e. internal hires) earn a wage premium over workers

who are hired externally, and the premium is increasing in the firm-specific experience of

redeployed workers. Models 1–4 of Table 3.7 compare the contractual wage of internal

and external hires, controlling for an establishment-occupation-year fixed effect.9 This

9The RAIS data provide information on several different measures of worker compensation,including the contractual wage and the actual amounts paid out to workers in any given year.While the two are highly correlated, for our analysis we use the contractual wage in order tonot capture any non-recurring payments that may otherwise be included in that year’s wagefor redeployed workers, such as relocation bonuses.

80

Table 3.7: Wages of Redeployed Workers

(1) (2) (3) (4) (5)

Redeploy 0.093∗∗∗ 0.032∗∗∗ 0.036∗∗∗ 0.095∗∗∗ 0.012(0.011) (0.009) (0.010) (0.015) (0.011)

Firm exp. 0.024∗∗∗ 0.024∗∗∗ 0.016∗∗∗(0.004) (0.004) (0.001)

Closing origin -0.041∗∗∗(0.014)

Redeploy ×Firm exp. 0.006∗∗

(0.003)Managers 0.014

(0.021)Professionals -0.011

(0.025)Technicians and Admin -0.008

(0.017)High school 0.016∗∗∗ 0.017∗∗∗ 0.017∗∗∗ 0.016∗∗∗ 0.041∗∗∗

(0.002) (0.002) (0.002) (0.002) (0.004)Higher Ed 0.174∗∗∗ 0.177∗∗∗ 0.177∗∗∗ 0.174∗∗∗ 0.252∗∗∗

(0.013) (0.012) (0.012) (0.013) (0.013)Age 0.004∗∗∗ 0.004∗∗∗ 0.004∗∗∗ 0.004∗∗∗ 0.011∗∗∗

(0.001) (0.001) (0.001) (0.001) (0.001)Age squared -0.000∗ -0.000∗∗ -0.000∗∗ -0.000∗ -0.000∗∗∗

(0.000) (0.000) (0.000) (0.000) (0.000)Female -0.021∗∗∗ -0.021∗∗∗ -0.021∗∗∗ -0.021∗∗∗ -0.047∗∗∗

(0.004) (0.004) (0.004) (0.004) (0.004)

Firms 5,991 5,991 5,991 5,991 6,749Observations 1,833,356 1,833,356 1,833,356 1,833,356 3,421,007

Note: All models include establishment-occupation-year fixed effects so that comparisons betweenredeployed workers and other workers are within establishment-occupation-year. The excluded cate-gory for the education dummies is “Less than high school”; the excluded category for the occupationcategories is “Service & Production” workers. Standard errors in parentheses are clustered by firm.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

81

model compares workers hired for the same occupation, in the same establishment, and

in the same year. On average, redeployed workers — i.e. those hired internally from

another unit of the same firm — earn about nine percent more than workers hired from

other firms (model 1). Model 2 shows that this wage premium is increasing in the firm-

specific work experience of redeployed workers. Specifically, a worker redeployed in their

first year with zero years of firm-specific experience earns an average wage premium of

3.5 percent over external hires. This premium increases by roughly 2.4 percent for each

year of experience working within the firm. Model 3 suggests that workers hired from

closing establishments within the same firm, however, earn much lower wage premiums

over external hires than workers moving from establishments that are not shutting down.

The comparison between workers moving from closing versus continuing establishments

should be interpreted cautiously; the use of establishment-occupation-year fixed effects

in the model means that this comparison depends on establishments hiring internal

workers from both closing and ongoing establishments to perform the same job in the

same year.10

One possible explanation for the large wage premium of internal over external hires

is unobservable differences in skill between those who are hired internally through rede-

ployment and workers hired through external labor markets. To address this possibility,

Model 5 of Table 3.7 compares the wages of redeployed workers and other workers

10There are 1,690 establishment-occupation-year cells with this variation (out of approximately1.8 million total observations.

82

Table 3.8: Adjustment Costs and Redeployment

(1) (2) (3) (4)

Diff. in growth 0.017 0.018(0.017) (0.017)

Industry similarity 0.029∗∗∗ 0.037∗∗∗(0.010) (0.011)

Log distance -0.125∗∗∗ -0.125∗∗∗(0.013) (0.013)

Dest. log employment 0.074∗∗∗ 0.075∗∗∗ 0.082∗∗∗ 0.083∗∗∗(0.011) (0.011) (0.010) (0.010)

Origin log employment 0.138∗∗∗ 0.138∗∗∗ 0.149∗∗∗ 0.150∗∗∗(0.024) (0.024) (0.023) (0.023)

Firms 2,286 2,286 2,282 2,282Observations 59,266 59,264 59,219 59,217

Note: Observations are establishment dyads (i.e. a destination and originestablishment pair) with positive redeployment. The dependent variable isthe natural logarithm of redeployments. All models include firm, destinationindustry, origin industry, and year fixed effects. Standard errors in parenthe-ses are clustered by firm.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

at the destination who were not hired that year. In other words, unlike models 1–4,

model 5 compares redeployed workers to their peers who were not redeployed. These

results show no statistically significant wage premium for redeployment in the absence

of firm-specific experience. This suggests that redeployed workers in fact resemble other

workers in their destination establishment. The interaction of redeployment with firm-

specific experience, however, indicates that redeployed workers may earn an additional

premium for firm-specific work experience. One additional year of firm-specific expe-

rience increases the wages of workers who are not redeployed by roughly 1.6 percent,

versus a 2.2 percent increase for those who are redeployed.

Firms redeploy workers more intensively between establishments in related industries

83

and establishments that are geographically closer to each other. Table 3.8 shows that a

one standard deviation increase in industry similarity is associated with a three percent

increase in redeployments. Models 3 and 4 show that a one percent increase in distance

between establishments is associated with 0.13 percent fewer redeployments. Together,

these results support the relatedness and distance hypotheses (i.e. hypotheses 3 and 4).

The results for the inducement hypothesis that favorable industry conditions in a desti-

nation establishment relative to an origin establishment are equivocal. The coefficients

in models 1 and 4 of Table 3.8 have the expected sign, but are not statistically significant.

Consistent with the results of Table 3.6 showing an increased likelihood of redeploy-

ment when closing an establishment, models 1–3 of Table 3.9 show that the intensity

of redeployment is also greater when closing an establishment even after controlling for

industry similarity, geographic distance, differences in industry growth, and establish-

ment size. Specifically, a closing origin establishment is associated with a roughly 50

percent increase in the number of worker redeployments (see model 1). Models 2–3,

however, fail to support the hypotheses that redeployments occurring when an estab-

lishment closes (i.e. under slack conditions) will be less sensitive to industry relatedness

and geographic distance. The coefficients on the interactions between an establishment

closing and industry similarity or geographic distance do not have the expected sign; the

positive and negative coefficients on these interactions respectively suggest that firms

may be more sensitive to industry relatedness and geographic distance when redeploy-

ing workers from a closing establishment. Some caution is warranted, however, when

84

Table 3.9: Adjustment Costs and Redeployment When Closing

(1) (2) (3)

Closing origin 0.508∗∗∗ 0.421∗∗∗ 0.728∗∗∗(0.081) (0.053) (0.106)

Closing origin ×Industry similarity 0.050∗

(0.030)Log distance -0.060∗∗∗

(0.018)Industry similarity 0.036∗∗∗ 0.032∗∗∗ 0.036∗∗∗

(0.010) (0.011) (0.010)Log distance -0.124∗∗∗ -0.124∗∗∗ -0.121∗∗∗

(0.013) (0.013) (0.013)Diff. in growth 0.014 0.015 0.014

(0.018) (0.018) (0.018)Dest. log employment 0.080∗∗∗ 0.081∗∗∗ 0.081∗∗∗

(0.010) (0.010) (0.010)Origin log employment 0.174∗∗∗ 0.173∗∗∗ 0.174∗∗∗

(0.020) (0.020) (0.020)

Firms 2,282 2,282 2,282Observations 59,217 59,217 59,217

Note: Observations are establishment dyads (i.e. a destination andorigin establishment pair) with positive redeployment. The dependentvariable is the natural logarithm of redeployments. All models includefirm, destination industry, origin industry, and year fixed effects. Stan-dard errors in parentheses are clustered by firm.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

85

interpreting these results because workers redeployed when establishments close may be

unlike workers redeployed during normal business conditions. Specifically, such workers

may be less adaptable to new industries or more sensitive to moving large geographic

distances.

3.5 Conclusion

This study has explored the extent and drivers of internal labor market activity in

multi-business firms in the context of Brazil. We have presented a simple framework

where two distinct forces can give rise to an incentive to redeploy workers: external

labor market frictions (hiring and firing costs) and workers’ possession of valuable, firm-

specific knowledge.

We find that Brazilian multi-business firms source a meaningful share of their workers

from within the firm. In an average year, the typical establishment sources 12.1 percent

of new hires internally. Studying what predicts whether a worker is redeployed, we find

evidence that both workers higher in the occupational hierarchy and workers with more

firm-specific experience are more likely to be redeployed. Managers, in particular, are

redeployed more than twice as often as the average worker.

We also study the wages of workers hired into a position through internal redeploy-

ment versus the external labor market. Comparing two workers hired into the same

narrow occupation, in the same establishment, in the same year, with otherwise similar

86

characteristics, we find that redeployed workers earn a nine percent wage premium over

those hired externally. On the other hand, redeployed workers do not earn a substantial

premium over otherwise comparable workers at the destination, suggesting that these

results are not driven by selection on unobservable worker quality (redeployed workers

being of better quality relative to other internal workers). The wage premium is consis-

tent with firm-specific experience rather than worker’s personal motivations or external

hiring frictions driving redeployment.

Our paper contributes to existing theory of resource redeployment, which has theo-

rized but rarely observed actual redeployment. Our results show that redeployment is

pervasive in the context of internal labor markets in multi-business firms, and most of-

ten does not involve the simultaneous exit of the origin business unit. Furthermore, our

paper contributes to the broader literature on internal labor markets. Compared to the

existing focus on vertical labor markets and horizontal labor markets as a response to

external labor market frictions, the results of our paper support the view that internal

labor markets also serve as conduits of firm-specific knowledge inside the firm.

Our results also provide directions for future research. One feature that we have

observed in the data is that redeployments are especially high when firms first open

new establishments. Tables 3.4 and 3.5 show that in such cases, 22.5 percent of all

initial workers and 42 percent of all managers of the new plants are sourced from other

units of the same firm. Understanding the strategies multi-business firms use when

they engage in “intrapreneurship”—in particular the type and nature of the human

87

resources allocated to new businesses—and whether the option to leverage their internal

labor markets provides a competitive advantage over independent startups constitutes

an important and interesting area for future research.

88

AConstruction of O∗NET Task Measures

This appendix lists the O∗NET scales used to construct the occupation task measures

used in Table 1.5. The O∗NET scales used in this paper are based on those in Acemoglu

and Autor (2011) and computer code from David Autor’s website.1

Acemoglu and Autor (2011) use two sub-measures of non-routine cognitive tasks:

“analytical” and “interpersonal.” For simplicity, I combine these two measures into a

single non-routine cognitive measure.

The computer code provided by David Autor for constructing task measures includes

two non-routine manual scales: “physical” and “interpersonal”. The interpersonal scale

is not used in Acemoglu and Autor (2011). I combine the two non-routine manual scales

1Available at https://economics.mit.edu/faculty/dautor/data/acemoglu (archived athttps://perma.cc/B7SK-VKUV).

89

in the computer code into a single non-routine manual measure.

Non-routine cognitive

4.A.2.a.4 Analyzing data/information

4.A.2.b.2 Thinking creatively

4.A.4.a.1 Interpreting information for others

4.A.4.a.4 Establishing and maintaining personal relationships

4.A.4.b.4 Guiding, directing and motivating subordinates

4.A.4.b.5 Coaching/developing others

Non-routine manual

4.A.3.a.4 Operating vehicles, mechanized devices, or equipment

4.C.2.d.1.g Spend time using hands to handle, control or feel objects, tools or

controls

1.A.2.a.2 Manual dexterity

1.A.1.f.1 Spatial orientation

2.B.1.a Social Perceptiveness

Routine cognitive

90

4.C.3.b.7 Importance of repeating the same tasks

4.C.3.b.4 Importance of being exact or accurate

4.C.3.b.8 Structured v. Unstructured work (reverse)

Routine manual

4.C.3.d.3 Pace determined by speed of equipment

4.A.3.a.3 Controlling machines and processes

4.C.2.d.1.i Spend time making repetitive motions

91

BVariable Definitions for Chapter 2

The variables for chapter 2—listed in Table 2.1—are defined as follows:

• Bills introduced is the number of bills introduced in the legislature.

• Laws enacted is the number of bills that became law.

• Tightening and Loosening laws are numbers of enacted laws that tightened and

loosened gun control respectively.

• Mass shooting is an indicator for state-years with a mass shooting in which three

or more people not romantically involved with or related to the shooter(s) were

killed.

• Fatalities is the total number of deaths in mass shootings in a state-year.

92

• Democratic and Republican Legislature are indicators for party control of the state

legislature.

• Republican governor is an indicator for Republican governors.

• Regular session indicates whether the legislature convened a regular (as opposed

to special) session to consider bills; some state legislatures only meet every other

year.

• Bill carryover is proportion of chambers in which bills are eligible for carryover

to the next session.

• Limited leg. topic is an indicator for legislative sessions during which bills are

limited to specific topics (e.g. appropriations).

• Legislature size is the number of lawmakers serving in the state legislature.

• The demographic controls are percentages of the state’s population, except for

Income per capita, which is measured in thousands of 1987 U.S. dollars.

93

CCoding Gun Laws

In order to facilitate accurate coding of gun legislation, coders were given a full manual

to explain the meaning of “tighten”, “loosen”, “neutral,” and “uncertain” along with

the examples in Table C.1. The table mimics the appearance of the Excel workbooks

used by the coders. The first bill creates a new crime related to firearms. It tightens

restrictions on firearms. The second bill makes it easier for people to acquire guns; it

loosens restrictions on firearms. The third bill is exclusively about parole officers; it

is neutral because it does not affect the general public. The fourth bill is uncertain

because the summary is a generic description that does not specify whether the law

tightens or loosens restrictions on firearms. The fifth bill both tightens and loosens; it

regulates gun shows, but also eliminates a restriction on firearm purchasers.

94

Table C.1: Coding Gun Laws

ID Summary Tighten Loosen Uncertain

1 Creates a new felony for firing a gun within 1,000 feet of aneducational facility.

1 0 0

2 Reduces the age limit for purchase of a handgun from 21 to18.

0 1 0

3 Allows parole officers to carry a loaded firearm while commut-ing to and from work.

0 0 0

4 Relates to the use of firearms in state parks and campgrounds. 0 0 15 Requires a license to operate a gun show. Eliminates the

waiting period for firearm sales if the purchaser has a validpermit to carry a concealed weapon.

1 1 0

Note: Table shows examples of coding laws based on bill summaries.

95

DEffect of Mass Shootings in Neighboring States

96

Tabl

eD

.1:

Mas

sSho

otin

gsin

Neig

hbor

ing

Stat

es,B

illsan

dLa

ws

Bills

Intro

duce

dLa

wsEn

acte

d

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Mas

ssh

ootin

g0.

147∗

∗0.

158∗

∗0.

097

0.08

7(0

.064

)(0

.073

)(0

.067

)(0

.080

)N

eigh

bor

shoo

ting

-0.0

520.

028

(0.0

41)

(0.0

48)

Cen

.div

ision

shoo

ting

-0.0

180.

020

(0.0

42)

(0.0

49)

Fata

litie

s0.

023∗

∗∗0.

022∗

∗∗0.

014∗

0.01

1(0

.007

)(0

.008

)(0

.008

)(0

.009

)N

eigh

bor

fata

litie

s-0

.002

0.00

2(0

.004

)(0

.004

)C

en.d

ivisi

onfa

talit

ies

0.00

20.

004

(0.0

05)

(0.0

04)

Polit

ical

Con

trol

s•

••

••

••

•In

stitu

tiona

lCon

trol

s•

••

••

••

•D

emog

raph

icC

ontr

ols

••

••

••

••

Stat

eFi

xed

Effec

ts•

••

••

••

•Ye

arFi

xed

Effec

ts•

••

••

••

•N

1,25

01,

250

1,25

01,

250

1,25

01,

250

1,25

01,

250

Not

e:T

hede

pend

ent

varia

ble

isth

enu

mbe

rof

firea

rm-r

elat

edbi

llsin

trod

uced

(mod

els

1–4)

orla

ws

enac

ted

(mod

els

5–8)

inst

ate

legi

slatu

res.

Nei

ghbo

rre

fers

tost

ates

with

ash

ared

bord

er;

Cen

.di

visi

onre

fers

tost

ates

with

inth

esa

me

Cen

sus

divi

sion.

Rob

ust

stan

dard

erro

rscl

uste

red

byst

ate

inpa

rent

hese

s.∗p<

.10,∗

∗p<

.05,

∗∗∗p<

.01

97

Table D.2: Mass Shootings in Neighboring States and Directions of PolicyChange

Tightening Laws Loosening Laws

(1) (2) (3) (4)

Shooting ×Republican legislature -0.015 -0.104 0.732∗∗∗ 0.809∗∗∗

(0.223) (0.260) (0.257) (0.275)Democratic legislature 0.072 0.082 -0.294 -0.169

(0.129) (0.122) (0.379) (0.416)Split legislature -0.242 -0.057 0.174 0.090

(0.249) (0.224) (0.331) (0.383)Neighbor shooting ×Republican legislature 0.208 -0.059

(0.135) (0.236)Democratic legislature -0.315∗∗ 0.370

(0.131) (0.246)Split legislature -0.157 -0.095

(0.202) (0.234)Cen. division shooting ×Republican legislature 0.166 -0.153

(0.154) (0.232)Democratic legislature -0.071 -0.141

(0.112) (0.175)Split legislature -0.258 0.146

(0.170) (0.258)

Political Controls • • • •Institutional Controls • • • •Demographic Controls • • • •State Fixed Effects • • • •Year Fixed Effects • • • •N 1,250 1,250 1,175 1,175

Note: The dependent variable is the number of firearm-related laws enacted thatmake gun laws stricter (models 1 and 2) or less strict (models 3 and 4). Neighborrefers to states with a shared border; Cen. division refers to states within thesame Census division. Robust standard errors clustered by state in parentheses.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

98

Table D.3: Fatalities in Neighboring States and Directions of Policy Change

Tightening Laws Loosening Laws

(1) (2) (3) (4)

Fatalities ×Republican legislature 0.014 0.002 0.151∗∗∗ 0.150∗∗∗

(0.047) (0.049) (0.031) (0.043)Democratic legislature 0.016 0.023 -0.046 -0.018

(0.014) (0.014) (0.053) (0.054)Split legislature 0.013 0.029∗∗ -0.022 -0.023

(0.012) (0.014) (0.019) (0.029)Neighbor fatalities ×Republican legislature 0.017 -0.005

(0.013) (0.022)Democratic legislature -0.025∗∗ 0.015

(0.012) (0.017)Split legislature -0.008 -0.035∗∗

(0.022) (0.018)Cen. division fatalities ×Republican legislature 0.020∗ 0.002

(0.011) (0.029)Democratic legislature -0.009 -0.031∗∗

(0.008) (0.015)Split legislature -0.015 0.007

(0.021) (0.023)

Political Controls • • • •Institutional Controls • • • •Demographic Controls • • • •State Fixed Effects • • • •Year Fixed Effects • • • •N 1,250 1,250 1,175 1,175

Note: The dependent variable is the number of firearm-related laws enacted thatmake gun laws stricter (models 1 and 2) or less strict (models 3 and 4). Neighborrefers to states with a shared border; Cen. division refers to states within thesame Census division. Robust standard errors clustered by state in parentheses.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

99

EEffect of Mass Shootings on Enacted Laws

100

Tabl

eE.

1:Eff

ecto

fMas

sSho

otin

gson

Enac

ted

Laws

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Mas

ssh

ootin

g0.

074

0.02

00.

079

0.09

8(0

.118

)(0

.108

)(0

.075

)(0

.067

)Fa

talit

ies

0.01

10.

007

0.01

10.

014∗

(0.0

13)

(0.0

11)

(0.0

08)

(0.0

08)

Reg

ular

sess

ion

3.39

3∗∗∗

3.38

4∗∗∗

3.38

9∗∗∗

3.37

8∗∗∗

(0.7

72)

(0.7

82)

(0.7

80)

(0.7

91)

Bill

carr

yove

r-0

.074

-0.0

79-0

.081

-0.0

88(0

.096

)(0

.095

)(0

.096

)(0

.095

)Li

mite

dle

g.to

pic

-0.6

02∗∗

-0.6

16∗∗

-0.5

97∗∗

-0.6

09∗∗

(0.2

60)

(0.2

66)

(0.2

58)

(0.2

65)

Legi

slatu

resiz

e-0

.012

∗∗∗

-0.0

07-0

.012

∗∗∗

-0.0

07(0

.004

)(0

.004

)(0

.004

)(0

.005

)D

em.l

egisl

atur

e-0

.023

-0.0

20(0

.082

)(0

.085

)R

ep.l

egisl

atur

e0.

269∗

∗∗0.

275∗

∗∗

(0.0

87)

(0.0

87)

Rep

.gov

erno

r-0

.002

-0.0

00(0

.051

)(0

.050

)

Dem

ogra

phic

Con

trol

s•

••

•St

ate

Fixe

dEff

ects

••

••

••

••

Year

Fixe

dEff

ects

••

••

••

N1,

250

1,25

01,

250

1,25

01,

250

1,25

01,

250

1,25

0

Not

e:T

hede

pend

entv

aria

ble

isth

enu

mbe

roffi

rear

m-r

elat

edla

wse

nact

edby

the

stat

e.R

obus

tsta

ndar

der

rors

clus

tere

dby

stat

ein

pare

nthe

ses.

∗p<

.10,∗

∗p<

.05,

∗∗∗p<

.01

101

FPredicting Mass Shootings

Tables F.1 and F.2 show the results of trying to predict mass shootings with demographic

and gun policy variables. The policy variables in Table F.2 are defined as follows:

• Handgun waiting period is the number of days purchasers must wait before ac-

cepting delivery of a handgun.

• Long-gun waiting period is similarly defined for long-guns (e.g. rifles and shot-

guns).

• Age 18+ transaction is an indicator for laws that prevent vendors from selling

handguns to minors or prevent minors from purchasing handguns.

• Age 21+ transaction is defined the same way for persons under 21.

102

• Handgun Permit System is an indicator for states that require permits to purchase

a handgun.

• Background check, all handgun sales is an indicator for requiring a background

check for all handgun transactions (including private sales).

• Background check, all firearm sales is an indicator for requiring a background

check for all firearm transactions (including private sales).

• Assault weapons ban is an indicator for states that ban some types of assault rifles

or pistols.

• Shall issue concealed carry is an indicator for states that require the permitting au-

thority to grant a license to anyone meeting the minimum statutory qualifications

(i.e. do not permit law enforcement discretion in issuing permits).

• No permit needed concealed carry is an indicator for states that allow concealed

carry without a permit.

103

Table F.1: Linear Probability Model for Mass Shootings Using Control Variables

(1) (2) (3) (4) (5) (6) (7)Lag bills introduced 0.000

(0.001)Lag laws enacted 0.000

(0.005)Lag tightening laws -0.000

(0.009)Lag loosening laws 0.016

(0.014)Dem. legislature 0.006 0.011 0.010 0.010 0.010 0.012

(0.044) (0.043) (0.045) (0.046) (0.045) (0.045)Rep. legislature -0.030 -0.030 -0.027 -0.027 -0.027 -0.028

(0.036) (0.035) (0.037) (0.037) (0.037) (0.038)Rep. governor -0.012 -0.010 -0.010 -0.010 -0.010 -0.010

(0.020) (0.020) (0.021) (0.021) (0.022) (0.021)Regular session 0.121 0.122∗ 0.122∗ 0.121 0.128∗

(0.077) (0.072) (0.067) (0.073) (0.073)Bill carryovera 0.057∗∗ 0.056∗ 0.055∗∗ 0.055∗∗ 0.056∗∗

(0.026) (0.029) (0.027) (0.027) (0.027)Limited leg. topic -0.044 -0.062 -0.062 -0.062 -0.065

(0.062) (0.059) (0.060) (0.059) (0.060)Legislature size 0.002 0.002 0.002 0.002 0.002

(0.001) (0.002) (0.002) (0.002) (0.002)Log population -0.191 -0.131 -0.140 -0.210 -0.211 -0.210 -0.210

(0.279) (0.265) (0.273) (0.297) (0.298) (0.297) (0.299)Elderly 0.001 -0.001 -0.002 0.004 0.004 0.004 0.003

(0.025) (0.024) (0.024) (0.024) (0.024) (0.025) (0.025)Under 25 -0.002 0.001 0.001 0.002 0.003 0.003 0.002

(0.020) (0.019) (0.019) (0.020) (0.020) (0.020) (0.020)Black -0.009 -0.009 -0.008 0.000 -0.000 -0.000 0.002

(0.015) (0.015) (0.015) (0.017) (0.018) (0.018) (0.017)Hispanic -0.001 -0.003 -0.002 -0.001 -0.001 -0.001 -0.001

(0.015) (0.015) (0.015) (0.016) (0.016) (0.016) (0.016)Unemployment 0.025∗∗ 0.025∗∗ 0.025∗∗ 0.024∗∗ 0.024∗∗ 0.024∗∗ 0.024∗∗

(0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.012)Income 0.013 0.012 0.013 0.013 0.013 0.013 0.013

(0.013) (0.013) (0.014) (0.014) (0.014) (0.014) (0.014)High school -0.004 -0.003 -0.002 -0.003 -0.003 -0.003 -0.003

(0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006)Veteran -0.003 -0.001 -0.001 -0.004 -0.004 -0.004 -0.004

(0.012) (0.012) (0.012) (0.013) (0.013) (0.012) (0.012)Divorced -0.004 -0.003 -0.002 0.001 0.001 0.001 0.001

(0.010) (0.010) (0.011) (0.011) (0.011) (0.011) (0.011)N 1,250 1,250 1,250 1,200 1,200 1,200 1,200

a There is no a priori reason to think bill carryover would be related to mass shootings; this correlationis insignificant when Virginia, which unlike most states, allows carryover in even years is dropped fromthe sample. Four of Virginia’s six mass shootings happened in even years.Note: All models include state and year fixed effects. Standard errors are clustered by state.∗ p < .10, ∗∗ p < .05

104

Table F.2: Linear Probability Model for Mass Shootings Using Policy Vari-ables

(1) (2)

Handgun waiting period (days) 0.002 0.002(0.005) (0.005)

Long-gun waiting period (days) -0.005 -0.005(0.019) (0.020)

Age 18+ for transaction 0.031 0.026(0.026) (0.028)

Age 21+ for transaction -0.067 -0.080(0.057) (0.056)

Handgun permit system -0.137 -0.136(0.094) (0.100)

Background check, all handgun sales -0.066 -0.066(0.080) (0.085)

Background check, all firearm sales 0.019 -0.005(0.116) (0.127)

Assault weapons ban 0.042 0.050(0.051) (0.053)

Shall issue concealed carry -0.010 -0.006(0.038) (0.039)

No permit needed concealed carry 0.162 0.181(0.188) (0.184)

Log population -0.494 -0.445(0.325) (0.310)

Political Controls •Demographic Controls • •N 1,250 1,250

Note: All models include state and year fixed effects. Standard errors are clus-tered by state.

105

GMass Shootings and State-Specific Time Trends

106

Table G.1: Effect of Mass Shootings on Gun Bill Introductions

(1) (2) (3) (4) (5) (6)

Mass shooting 0.074 0.151∗∗ 0.157∗∗(0.075) (0.069) (0.063)

Fatalities 0.020∗∗ 0.024∗∗∗ 0.024∗∗∗(0.010) (0.008) (0.007)

Institutional Controls • • • •Political Controls • •Demographic Controls • •State Fixed Effects • • • • • •State-Specific Trends • • • • • •Year Fixed Effects • • • • • •N 1,250 1,250 1,250 1,250 1,250 1,250

Note: The dependent variable is the number of firearm-related bills introduced in state legisla-tures. Robust standard errors clustered by state in parentheses. Variables are identical to thosein Table 2.2.∗∗ p < .05, ∗∗∗ p < .01

Table G.2: Mass Shootings, Ordinary Gun Homicides, and Bill Intro-ductions

(1) (2) (3)

Mass shooting fatalities / 100,000 1.504∗∗∗ 1.481∗∗∗ 1.409∗∗∗(0.323) (0.261) (0.195)

Ordinary gun homicides / 100,000 0.010 0.005 0.007(0.058) (0.055) (0.049)

Institutional Controls • •Political Controls •Demographic Controls •State Fixed Effects • • •State-Specific Trends • • •Year Fixed Effects • • •N 1,250 1,250 1,250

Note: The dependent variable is the number of firearm-related bills in-troduced in state legislatures. Robust standard errors clustered by state inparentheses. Variables are identical to those in Table 2.3.∗∗∗ p < .01

107

HPlacebo Mass Shooting Analyses

We randomly assign placebo mass shootings to state-years in which no actual shooting

occurred with probability equal to each state’s frequency of shootings, and randomly

draw a fatality count from the empirical distribution of fatalities. We then re-run the

models and calculate the test statistic for the placebo shooting and fatality coefficients.

The percentiles in Tables H.1 and H.2 are based on 1,000 replications. The “pooled”

rows in Table H.2 mirror the models in Table 2.4 without interaction effects.

108

Table H.1: Placebo Analysis for Bill Introductions

Percentiles of Placebo Test Statistic

Actual 1st 5th 10th 90th 95th 99th

Shooting Indicator (model 4) 2.37 -3.69 -2.80 -2.33 0.71 1.06 1.90Shooting Fatalities (model 8) 3.29 -4.01 -2.81 -2.40 0.94 1.35 2.32

Note: Models mirror those of Table 2.2.

Table H.2: Placebo Analysis for Enacted Laws

Percentiles of Placebo Test Statistic

Actual 1st 5th 10th 90th 95th 99th

Tightening LawsPooled shooting -0.30 -2.19 -1.40 -1.04 1.82 2.29 3.02Pooled fatalities 1.88 -2.39 -1.64 -1.24 1.80 2.28 3.44Shooting ×Republican legislature -0.07 -2.93 -1.71 -1.18 1.63 2.18 3.28Democratic legislature 0.26 -2.47 -1.63 -1.33 1.45 1.96 2.60Split legislature -0.82 -2.65 -1.46 -0.98 2.13 2.60 4.14

Fatalities ×Republican legislature 0.36 -2.80 -1.86 -1.33 1.68 2.46 3.72Democratic legislature 0.93 -2.81 -1.80 -1.47 1.52 2.02 3.49Split legislature 1.15 -2.79 -1.65 -1.09 2.26 2.94 4.82

Loosening LawsPooled shooting 1.36 -3.19 -2.28 -1.89 0.84 1.26 2.11Pooled fatalities 0.36 -3.04 -2.25 -1.82 0.93 1.46 2.58Shooting ×Republican legislature 2.87 -2.98 -2.27 -1.97 0.68 1.05 1.69Democratic legislature -0.62 -2.69 -1.79 -1.40 1.23 1.66 2.80Split legislature 0.49 -3.10 -2.12 -1.70 1.40 1.86 2.89

Fatalities ×Republican legislature 4.90 -2.80 -2.24 -1.89 0.93 1.35 2.22Democratic legislature -0.87 -2.89 -1.90 -1.48 1.47 1.94 2.87Split legislature -1.00 -3.22 -2.12 -1.75 1.47 2.04 3.32

Note: Models mirror those of Table 2.4. The models for tightening laws mirrormodels 1–4; models for loosening laws mirror models 5–8.

109

IExcluding States from Mass Shooting Analyses

These analyses exclude each state, one at a time, from our sample. Each graph plots

the resulting 50 regression coefficients (from smallest to largest) along with a 95 percent

confidence interval and estimates using the full sample of all states. The state abbrevi-

ations in the figures indicate the state that was dropped from the sample and mark the

resulting point estimate. Vertical bars represent 95 percent confidence intervals. The

solid, horizontal line indicates the point estimate from the complete sample (presented

in chapter 2), and dotted, horizontal lines represent the lower and upper bounds of the

95 percent confidence interval for the full sample estimate. Removing individual states

has little effect on the coefficient estimates, supporting the claim that the effect of mass

shootings on gun policy is not driven by an individual state or shooting.

110

Figure I.1: Effect of Mass Shooting on Bill Introductions

ALAKAZ ARCA COCT

DE FLGAHI ID ILIN IAKS KY LAME

MDMA

MIMNMSMO MT NENV NHNJ NM

NY

NC ND OH OKORPA

RISC

SD

TN

TX UT VTVAWA

WV WIWY

0.00

0.05

0.10

0.15

0.20

0.25

0.30

Coe

ffici

ent E

stim

ate

Figure I.2: Effect of Mass Shooting on Laws Enacted

AL

AK AZ

ARCA CO

CT DEFL GA HI ID

IL

IN IAKS KYLA

MEMDMA

MI

MNMS MOMT NE NVNH NJ NM

NY

NCND OH

OKOR PARISC SD

TN

TX

UT VT

VA

WAWV WI WY

−0.10

−0.05

0.00

0.05

0.10

0.15

0.20

0.25

Coe

ffici

ent E

stim

ate

111

Figure I.3: Effect of Republican Legislature Mass Shooting on Loosening Laws Enacted

ALAK AZ

ARCA

COCT DE

FL

GAHIID

ILINIA

KS KYLAMEMD MAMI MN MSMO MT NE

NV

NH NJNMNY

NC

ND

OH

OKOR PA

RI SCSD

TN

TX

UT

VTVA

WA WV

WI

WY

0.0

0.2

0.4

0.6

0.8

1.0

1.2

Coe

ffici

ent E

stim

ate

Figure I.4: Effect of Democratic Legislature Mass Shooting on Loosening Laws Enacted

AL

AK AZ

AR

CA

COCTDE FLGA HIIDIL

IN IA KSKYLA

MEMDMAMI MNMS MOMT NE NVNHNJ NM

NYNCND OH OK ORPARI

SCSD TNTX UTVT VA WAWVWI WY

−1.5

−1.0

−0.5

0.0

0.5

Coe

ffici

ent E

stim

ate

112

Figure I.5: Effect of Republican Legislature Mass Shooting on Tightening Laws Enacted

AL AK

AZ

AR

CACO CT DE

FL

GAHI IDIL INIA

KS KYLA

MEMD MA

MI

MN MSMOMTNENV NHNJ NMNY NC

ND

OHOKOR PA

RISCSD

TN

TXUTVT

VAWAWV

WIWY

−0.6

−0.4

−0.2

0.0

0.2

0.4

0.6

Coe

ffici

ent E

stim

ate

Figure I.6: Effect of Democratic Legislature Mass Shooting on Tightening Laws Enacted

ALAKAZ

AR

CA

COCT

DE

FLGA

HI IDIL

IN IAKSKY

LA

ME

MDMA MI MNMS MOMT NE NVNHNJ

NM NYNC

NDOH OKOR PARI SC

SDTN TXUTVT

VA

WAWV

WIWY

−0.3

−0.2

−0.1

0.0

0.1

0.2

0.3

0.4

0.5

Coe

ffici

ent E

stim

ate

113

JGun Ownership, Shootings, and Enacted Laws

Table J.1 adds a proxy for gun ownership—the percentage of suicides that are firearm

related (Cook and Ludwig, 2006)—to the analysis of tightening and loosening laws

presented in Table 2.4. Models 1 and 4 of Table J.1 show that the main results do not

change when adding this control variable. The other models suggest that the respond

of Republican legislatures cannot be explained by rated of gun ownership.

114

Table J.1: Effects of Mass Shootings with Gun Ownership Proxy

Tightening Laws Loosening Laws

(1) (2) (3) (4) (5) (6)

Mass shooting -0.209 -0.469 -0.325 0.189 0.720 0.527(0.260) (0.584) (0.511) (0.347) (0.786) (0.710)

Shooting ×Rep. legislature 0.193 0.169 0.533 0.580

(0.399) (0.385) (0.484) (0.482)Dem. legislature 0.255 0.243 -0.385 -0.382

(0.308) (0.295) (0.467) (0.440)Gun suicide percent 0.005 0.005 -0.010 -0.005

(0.009) (0.010) (0.014) (0.014)Gun suicide percent 0.037 0.036 0.035 0.106∗ 0.109∗ 0.116∗

(0.027) (0.027) (0.027) (0.060) (0.061) (0.061)Dem. legislature 0.062 0.062 0.100 -0.255 -0.261 -0.322∗

(0.171) (0.170) (0.149) (0.229) (0.225) (0.194)Rep. legislature 0.145 0.149 0.177 0.423∗∗ 0.413∗ 0.510∗∗∗

(0.143) (0.143) (0.133) (0.216) (0.218) (0.192)Rep. governor -0.040 -0.046 -0.046 -0.092 -0.082 -0.108

(0.086) (0.088) (0.089) (0.164) (0.162) (0.166)

Institutional Controls • • • • • •Demographic Controls • • • • • •State Fixed Effects • • • • • •Year Fixed Effects • • • • • •N 1,250 1,250 1,250 1,175 1,175 1,175

Note: The dependent variable is the number of firearm-related laws enacted that makegun laws stricter (models 1–3) or less strict (models 4–6). Gun suicide percent is the five-year moving average of the percentage of suicides that are firearm-related and is used toproxy for gun ownership (Cook and Ludwig, 2006). Robust standard errors clustered bystate in parentheses.∗ p < .10, ∗∗ p < .05, ∗∗∗ p < .01

115

KMass Shootings as an Instrument for Gun Policy

In this appendix we use mass shootings as an instrumental variable to study the impact

of gun laws on gun deaths. We start with the following model:

lnDst = αs + θt + βGun Controlst + δ′Zst + ϵst

where Dst is non-mass shooting gun deaths per 100,000 people in state s and year t,

αs and θt are state and year fixed effects, Gun Controlst is an index representing the

strictness of gun policy, and Zst is a vector of controls—demographic, political, and

economic factors-–-that potentially affect gun deaths. We use the same variables as

Levitt (1996) as controls, but also include dummies for Republican and Democratic

trifectas or legislatures, and a dummy for Republican governors.

We do not directly observe Gun Controlst; instead, we observe the enactment of new

116

laws that change gun policy. Therefore, we estimate the equation in first differences:

∆ lnDst = λt + βNew Gun Lawsst +∆Zstδ +∆ϵst

where New Gun Lawsst −∆Gun Controlst is negative for laws that loosen gun control

and positive for laws that tighten gun control (according to our coders, see data de-

scription). Based on our main results, we instrument for gun laws using the first lags

of mass shooting fatalities and the interaction of lagged mass shooting fatalities with

Republican control of state government. The former should be positively correlated

with new laws and the latter negatively correlated with new laws.

We estimate the model using Fuller’s (1977) modified LIML with α = 1 (Baum,

Schaffer, and Stillman, 2007). First stage results suggest the instruments are weak

despite being jointly significant (F = 5.98) with the expected sign (Stock and Yogo,

2005). The coefficients on the exogenous instruments in the reduced form equation

for firearm deaths are not significant, but also have the expected signs (negative for

lagged mass shooting fatalities and positive for the interaction of lagged fatalities with

Republican control of government). Our estimate β̂ is −0.016 with standard error

0.013. A conditional likelihood ratio test (Moreira, 2003; Andrews, Moreira, and Stock,

2006; Finlay and Magnusson, 2009) cannot reject the null hypothesis that β = 0 (p =

0.24). We also estimated models that include proxies for gun ownership. Including the

percentage of suicides committed with a gun (Cook and Ludwig, 2006) does not change

our inference.

117

References

Abrams, David S. 2012. “Estimating the Deterrent Effect of Incarceration UsingSentencing Enhancements.” American Economic Journal: Applied Economics4 (4):32–56.

Acemoglu, Daron and David Autor. 2011. “Skills, Tasks and Technologies:Implications for Employment and Earnings.” In Handbook of Labor Economics, vol. 4.Elsevier, 1043–1171.

Akerman, Anders, Ingvil Gaarder, and Magne Mogstad. 2015. “The SkillComplementarity of Broadband Internet.” The Quarterly Journal of Economics130 (4):1781–1824.

Almeida, Rita K., Carlos H.L. Corseuil, and Jennifer P. Poole. 2017. “The Impact ofDigital Technologies on Routine Tasks: Do Labor Policies Matter?” Policy ResearchWorking Paper WPS8187, The World Bank.

Andrews, Donald W.K., Marcelo J. Moreira, and James H. Stock. 2006. “OptimalTwo-Sided Invariant Similar Tests for Instrumental Variables Regression.”Econometrica 74 (3):715–752.

Autor, David. 2013. “The ‘Task Approach’ to Labor Markets: An Overview.”Journal for Labour Market Research 46 (3):185–199.

Autor, David, Frank Levy, and Richard J. Murnane. 2003. “The Skill Content ofRecent Technological Change: An Empirical Exploration.” The Quarterly Journal ofEconomics 118 (4):1279–1333.

Bardhan, Pranab and Dilip Mookherjee. 2010. “Determinants of RedistributivePolitics: An Empirical Analysis of Land Reforms in West Bengal, India.” AmericanEconomic Review 100 (4):1572–1600.

118

Barney, Jay. 1991. “Firm Resources and Sustained Competitive Advantage.” Journalof Management 17 (1):99–120.

Baum, Christopher F., Mark E. Schaffer, and Steven Stillman. 2007. “EnhancedRoutines for Instrumental Variables/Generalized Method of Moments Estimationand Testing.” The Stata Journal 7 (4):465––506.

Baumgartner, Frank R. and Bryan D. Jones. 1993. Agendas and Instability inAmerican Politics. University of Chicago Press, 1st ed.

Belenzon, Sharon and Ulya Tsolmon. 2016. “Market Frictions and the CompetitiveEffects of Internal Labor Markets.” Strategic Management Journal 37 (2):315–334.

Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. 2004. “How MuchShould We Trust Differences-in-Differences Estimates?” The Quarterly Journal ofEconomics 119 (1):249–275.

Bloom, Nicholas, Luis Garicano, Raffaella Sadun, and John Van Reenen. 2014. “TheDistinct Effects of Information Technology and Communication Technology on FirmOrganization.” Management Science 60 (12):2859–2885.

Bouton, Laurent, Paola Conconi, Francisco Pino, and Maurizio Zanardi. 2014. “Gunsand Votes.” Working Paper 20253, National Bureau of Economic Research. URLhttp://www.nber.org/papers/w20253.

Caliendo, Lorenzo, Ferdinando Monte, and Esteban Rossi-Hansberg. 2015. “TheAnatomy of French Production Hierarchies.” Journal of Political Economy124 (4):809–852.

Cestone, Giacinta, Giovanni Pica, Chiara Fumagalliz, and Francis Kramarz. 2017.“Insurance Between Firms: The Role of Internal Labor Markets.” ECGI WorkingPaper Series in Finance.

Chetverikov, Denis, Bradley Larsen, and Christopher Palmer. 2016. “IV QuantileRegression for Group-Level Treatments, with an Application to the DistributionalEffects of Trade.” Econometrica 84 (2):809–833.

Chin, M. K. and Matthew Semadeni. 2017. “CEO Political Ideologies and PayEgalitarianism Within Top Management Teams.” Strategic Management Journal38 (8):1608–1625.

119

Cobb, J. Adam. 2016. “How Firms Shape Income Inequality: Stakeholder Power,Executive Decision Making, and the Structuring of Employment Relationships.”Academy of Management Review 41 (2):324–348.

Cook, Philip J. and Jens Ludwig. 2006. “The Social Costs of Gun Ownership.”Journal of Public Economics 90 (1–2):379–391.

Doeringer, Peter B. and Michael J. Piore. 1971. Internal Labor Markets andManpower Analysis. Lexington, MA: Heath Lexington Books.

Donald, Stephen G. and Kevin Lang. 2007. “Inference with Difference-in-Differencesand Other Panel Data.” Review of Economics and Statistics 89 (2):221–233.

Duggan, Mark. 2001. “More Guns, More Crime.” Journal of Political Economy109 (5):1086–1114.

Dutz, Mark A., Lucas Ferreira Mation, Stephen D. O’Connell, and Robert D. Willig.2017. “Economywide and Sectoral Impacts on Workers of Brazil’s Internet Rollout.”Policy Research Working Paper WPS8042, The World Bank.

Duwe, Grant. 2007. Mass Murder in the United States: A History. McFarland &Company, Inc.

Egan, Patrick J. and Megan Mullin. 2012. “Turning Personal Experience intoPolitical Attitudes: The Effect of Local Weather on Americans’ Perceptions aboutGlobal Warming.” The Journal of Politics 74 (3):796–809.

Eisensee, Thomas and David Strömberg. 2007. “News Droughts, News Floods, andU.S. Disaster Relief.” The Quarterly Journal of Economics 122 (2):693–728.

Fama, Eugene F. 1980. “Agency Problems and the Theory of the Firm.” Journal ofPolitical Economy 88 (2):288–307.

Finlay, Keith and Leandro M. Magnusson. 2009. “Implementing Weak-InstrumentRobust Tests for a General Class of Instrumental-Variables Models.” The StataJournal 9 (3):398––421.

Folta, Timothy B, Constance E Helfat, and Samina Karim. 2016. “ExaminingResource Redeployment in Multi-Business Firms.” In Resource Redeployment and

120

Corporate Strategy, Advances in Strategic Management, vol. 35, edited by Timothy BFolta, Constance E Helfat, and Samina Karim. Bingley, UK: Emerald GroupPublishing Limited, 1–17.

Forman, Chris, Avi Goldfarb, and Shane Greenstein. 2012. “The Internet and LocalWages: A Puzzle.” American Economic Review 102 (1):556––75.

Fredrickson, James W., Alison Davis-Blake, and WM. Gerard Sanders. 2010.“Sharing the Wealth: Social Comparisons and Pay Dispersion in the CEO’s TopTeam.” Strategic Management Journal 31 (10):1031–1053.

Fuller, Wayne A. 1977. “Some Properties of a Modification of the LimitedInformation Estimator.” Econometrica 45 (4):939–953.

Gartenberg, Claudine and Julie Wulf. 2017a. “Islands of Equality: Competition andPay Inequality Within and Across Firm Boundaries.” Unpublished.

———. 2017b. “Pay Harmony? Social Comparison and Performance Compensationin Multibusiness Firms.” Organization Science 28 (1):39–55.

Gillett, Sharon E., William H. Lehr, Carlos A. Osorio, and Marvin A. Sirbu. 2006.“Measuring Broadband’s Economic Impact.” National Technical Assistance, Training,Research, and Evaluation Project 99-07-13829, U.S. Department of Commerce.

Godwin, Marcia L. and Jean Reith Schroedel. 1998. “Gun Control Politics inCalifornia.” In The Changing Politics of Gun Control, edited by John M. Bruce andClyde Wilcox. Rowman & Littlefield Publishers, Inc., 88–110.

Goss, Kristin A. 2006. Disarmed: The Missing Movement for Gun Control inAmerica. Princeton Studies in American Politics: Historical, International, andComparative Perspectives. Princeton University Press.

Grant, Robert M. 1996. “Toward a Knowledge-Based Theory of the Firm.” StrategicManagement Journal 17:109–122.

Groysberg, Boris, Linda-Eling Lee, and Ashish Nanda. 2008. “Can They Take It withThem? The Portability of Star Knowledge Workers’ Performance.” ManagementScience 54 (7):1213–1230.

121

Helfat, Constance E and Kathleen M Eisenhardt. 2004. “Inter-Temporal Economiesof Scope, Organizational Modularity, and the Dynamics of Diversification.” StrategicManagement Journal 25 (13):1217–1232.

Hjort, Jonas and Jonas Poulsen. 2017. “The Arrival of Fast Internet and Employmentin Africa.” Working Paper 23582, National Bureau of Economic Research.

Jung, Juan and Enrique López-Bazo. 2017. “On the Regional Impact of Broadbandon Productivity: The Case of Brazil.” Working Paper 2017/08, Research Institute ofApplied Economics.

Kacperczyk, Aleksandra and Chanchal Balachandran. 2018. “Vertical and HorizontalWage Dispersion and Mobility Outcomes: Evidence from the Swedish Microdata.”Organization Science 29 (1):17–38.

Karim, Samina and Will Mitchell. 2004. “Innovating Through Acquisition andInternal Development: A Quarter-Century of Boundary Evolution at Johnson &Johnson.” Long Range Planning 37 (6):525–547.

Khanna, Tarun and Krishna Palepu. 2000. “Is Group Affiliation Profitable inEmerging Markets? An Analysis of Diversified Indian Business Groups.” The Journalof Finance 55 (2):867–891.

Kingdon, John W. 1984. Agendas, Alternatives, and Public Policies. Little, Brown.

Kogut, Bruce and Udo Zander. 1992. “Knowledge of the Firm, CombinativeCapabilities, and the Replication of Technology.” Organization Science 3 (3):383–397.

Kolko, Jed. 2012. “Broadband and Local Growth.” Journal of Urban Economics71 (1):100–113.

Krouse, William and Daniel J. Richardson. 2015. “Mass Murder with Firearms:Incidents and Victims, 1999–2013.” CRS Report 7-5700, Congressional ResearchService.

Lafontaine, Francine and Jagadeesh Sivadasan. 2009. “Do Labor Market RigiditiesHave Microeconomic Effects?” American Economic Journal: Applied Economics1 (2):88–127.

122

Lamont, Owen. 1997. “Cash Flow and Investment: Evidence from Internal CapitalMarkets.” The Journal of Finance 52 (1):83–109.

Larkin, Ian, Lamar Pierce, and Francesca Gino. 2012. “The Psychological Costs ofPay-for-Performance: Implications for the Strategic Compensation of Employees.”Strategic Management Journal 33 (10):1194–1214.

Levinthal, Daniel A and Brian Wu. 2010. “Opportunity Costs and Non-Scale FreeCapabilities: Profit Maximization, Corporate Scope, and Profit Margins.” StrategicManagement Journal 31 (7):780–801.

Levitt, Steven D. 1996. “The Effect of Prison Population Size on Crime Rates:Evidence from Prison Overcrowding Litigation.” The Quarterly Journal of Economics111 (2):319–351.

Lieberman, Marvin, Gwendolyn Lee, and Timothy Folta. 2016. “Entry, Exit, and thePotential for Resource Redeployment.” Strategic Management Journal 38 (3):526–544.

Ludwig, Jens and Philip J. Cook. 2000. “Homicide and Suicide Rates Associatedwith Implementation of the Brady Handgun Violence Prevention Act.” JAMA284 (5):585–591.

Ludwig, Jens and Philip J Cook, editors. 2003. Evaluating Gun Policy: Effects onCrime and Violence. Brookings Institution Press.

Makowsky, Michael D. and Thomas Stratmann. 2009. “Political Economy at AnySpeed: What Determines Traffic Citations?” The American Economic Review99 (1):509–527.

Mawdsley, John K. and Deepak Somaya. 2016. “Employee Mobility andOrganizational Outcomes: An Integrative Conceptual Framework and ResearchAgenda.” Journal of Management 42 (1):85–113.

Moreira, Marcelo J. 2003. “A Conditional Likelihood Ratio Test for StructuralModels.” Econometrica 71 (4):1027–1048.

Morris, Shad S., Sharon A. Alvarez, Jay B. Barney, and Janice C. Molloy. 2017.“Firm-Specific Human Capital Investments as a Signal of General Value: Revisiting

123

Assumptions about Human Capital and How it is Managed.” Strategic ManagementJournal 38 (4):912–919.

Morton, Robert J. and Mark A. Hilts, editors. 2008. Serial Murder:Multi-disciplinary Perspectives for Investigators. Federal Bureau of Investigation.URL http://www.fbi.gov/stats-services/publications/serial-murder/serial-murder-july-2008-pdf. Serial Murder Symposium, San Antonio, TX,August 29–September 2, 2005.

Nickerson, Jack A. and Todd R. Zenger. 2008. “Envy, Comparison Costs, and theEconomic Theory of the Firm.” Strategic Management Journal 29 (13):1429–1449.

Oi, Walter Y. and Todd L. Idson. 1999. “Firm Size and Wages.” In Handbook ofLabor Economics, vol. 3. Elsevier, 2165–2214.

Penrose, Edith. 1959. The Theory of the Growth of the Firm. Oxford UniversityPress.

Sakhartov, Arkadiy V and Timothy Folta. 2014. “Resource Relatedness,Redeployability, and Firm Value.” Strategic Management Journal 35 (12):1781–1797.

Sakhartov, Arkadiy V. and Timothy B. Folta. 2015. “Getting Beyond Relatedness asa Driver of Corporate Value.” Strategic Management Journal 36 (13):1939–1959.

Schuman, Howard and Stanley Presser. 1981. “The Attitude-Action Connection andthe Issue of Gun Control.” The Annals of the American Academy of Political andSocial Science 455 (1):40–47.

Shin, Hyun-Han and René M. Stulz. 1998. “Are Internal Capital Markets Efficient?”The Quarterly Journal of Economics 113 (2):531–552.

Siegel, Phyllis A. and Donald C. Hambrick. 2005. “Pay Disparities Within TopManagement Groups: Evidence of Harmful Effects on Performance ofHigh-Technology Firms.” Organization Science 16 (3):259–274.

Spitzer, Robert J. 2012. The Politics of Gun Control. Paradigm Publishers.

Stanford Geospatial Center and Stanford Libraries. 2015. “Stanford Mass Shootingsin America.” URLhttps://library.stanford.edu/projects/mass-shootings-america.

124

Stein, Jeremy C. 1997. “Internal Capital Markets and the Competition for CorporateResources.” The Journal of Finance 52 (1):111–133.

Stock, James H. and Motohiro Yogo. 2005. “Identification and Inference forEconometric Models.” In In Identification and Inference for Econometric Models,edited by Donald W.K. Andrews and James H. Stock. Cambridge University Press,80–108.

Tate, Geoffrey and Liu Yang. 2015. “The Bright Side of Corporate Diversification:Evidence from Internal Labor Markets.” Review of Financial Studies28 (8):2203–2249.

Teece, David J. 1981. “The Market for Know-How and the Efficient InternationalTransfer of Technology.” The Annals of the American Academy of Political and SocialScience 458 (1):81–96.

Teece, David J. 1982. “Towards an Economic Theory of the Multiproduct Firm.”Journal of Economic Behavior & Organization 3 (1):39–63.

USA Today. 2013. “Behind the Bloodshed: The Untold Story of America’s MassKillings.” URLhttp://www.gannett-cdn.com/GDContent/mass-killings/index.html.

Wernerfelt, Birger. 1984. “A Resource-Based View of the Firm.” StrategicManagement Journal 5 (2):171–180.

125