Labor Laws and Innovation1

Viral V. Acharya

NYU-Stern, CEPR, ECGI and NBER

vacharya@stern.nyu.edu

Ramin P. Baghai

Stockholm School of Economics

ramin.baghai@hhs.se

Krishnamurthy V. Subramanian

Indian School of Business

krishnamurthy subramanian@isb.edu

June 2012

1We are grateful to Amit Seru and Vikrant Vig for helpful comments and to Anusha Chari, Rich Mathews,Amalia R. Miller, and Radha Iyengar for their insightful discussions. Furthermore, we would like to thankseminar and conference participants at the American Law and Economics Annual Meeting (2009), WesternFinance Association Annual Meeting (2009), NBER Summer Institutes on Innovation Policy and the Econ-omy (2009) and Law and Economics (2009), Summer Research Conference in Finance (2009) at the IndianSchool of Business (ISB), Cambridge University (Centre for Business Research), Emory University, LondonBusiness School, NYU Microeconomics seminar, and NYU Stern for valuable comments and suggestions. Wethank Hanh Le and Chandrasekhar Mangipudi for excellent research assistance.

Abstract

Labor Laws and Innovation

We show that dismissal laws – laws that prevent employers from arbitrarily discharging em-

ployees – spur firm-level innovation. Dismissal laws limit employers’ ability to hold up innovating

employees after an innovation is successful. By reducing the possibility of hold-up, these laws

enhance employees’ innovative efforts and encourage firms to invest in risky, but potentially mould-

breaking, projects. We formalize this intuition and provide supporting empirical evidence across two

different empirical settings that exploit complementary variation in dismissal laws: (i) country-level

changes in dismissal laws; and (ii) a regression-discontinuity design examining firm-level innovation

after the passage of the 1989 WARN Act in the U.S. As the centerpiece of our identification in the

cross-country tests, we control for all country-level variables using fixed effects for each country–

year pair and undertake triple-difference tests to find that the effect of dismissal laws is stronger in

more innovation-intensive sectors than in less innovation-intensive ones.

JEL: F30, G31, J5, J8, K31.

Keywords: Labor laws, R&D, Technological change, Law and finance, Entrepreneurship, Growth.

I Introduction

Do legal institutions of an economy affect the pattern of its real investments, and, in turn,

its economic growth? In this paper, we focus on one specific aspect of this overarching theme.

In particular, we investigate whether the legal framework governing the relationships between

employees and their employers affects the extent of innovation in an economy.

While the inefficiencies and rigidities associated with stringent labor laws — laws that prevent

employers from seamlessly negotiating and/or terminating labor contracts with employees — are

much celebrated in the academic literature1 and the media, this discussion is generally centered

around the ex post effects of labor laws.2 In particular, it is clear that once the situation to

renegotiate or terminate an employment contract has arisen, tying down an employer’s hands from

doing so can lead to ex post inefficient outcomes. Much less studied, however, is the ex ante

incentive effect of such strong labor laws. Might stringent labor laws – even if as an unintended

consequence – provide firms a commitment device to not punish short-run failures and to not

hold up their employees in case of successful innovations and thereby spur employees to undertake

activities that are value-maximizing in the long-run?

This question assumes importance particularly in light of recent evidence that laws and contracts

that exhibit tolerance to failure can be instrumental in fostering innovation and economic growth.

Acharya and Subramanian (2009) report that the ex post inefficient continuations engendered by

debtor-friendly bankruptcy laws encourage ex ante risk-taking and, thereby, promote firm-level

innovation and country-level economic growth. Manso (2011) shows theoretically that the optimal

contract to motivate innovation not only exhibits tolerance for short-term failure but also, in fact,

rewards interim failure to create the incentives for successful innovation in the long-term; Ederer

and Manso (2010) find evidence supporting this thesis. Further, Tian and Wang (2011) show using

a sample of venture capital backed IPO firms that tolerance for failure among venture capitalists

spurs innovation in their portfolio firms.

In this paper, we contribute to this recent literature by highlighting that dismissal laws –

laws that prevent firms from arbitrarily discharging employees – indeed appear to have an ex-ante

positive incentive effect by encouraging firms and their employees to engage in more successful, and

more significant, innovative pursuits. We develop a theoretical model that formalizes this effect

and provide empirical evidence using two different settings. First, we exploit country-level changes

in dismissal laws from 1970–2002 for four countries – U.S., U.K., France, and Germany. Second,

we employ a regression-discontinuity design to examine the effect of the passage of the WARN Act

in 1989 on firm-level innovation within the U.S.

Since innovation involves considerable exploration, the difficulty in describing innovative ac-

tivities ex ante, combined with the possibility of ex-post renegotiation, make it difficult to write

1Botero et al. (2004), for example, claim that heavier regulation of labor leads to adverse consequences for labormarket participation and unemployment.

2For example, strong labor market regulation is often blamed to be one of the reasons for Europe’s economicunder-performance compared to the U.S. For a study articulating this theme, see the study of France and Germanyby the McKinsey Global Institute (1997).

1

complete contracts in innovative settings (Aghion and Tirole, 1994). Therefore, to appropriate

a larger share of the substantial payoff from successful innovation, innovative firms may hold up

employees that contributed to a successful innovation. In fact, a recent high-profile court case filed

against the video-game company Activision by its former employees West and Zampella highlights

such possible hold-up.3 Dismissal laws can help to limit the occurrence of hold-up and thereby

increase the employee’s innovative effort. To formalize this intuition and to obtain sharp empirical

predictions, we develop a theoretical model that leads to the following hypotheses:

Hypothesis 1: Stronger dismissal laws lead to greater innovation.

Since the ex-ante incentive effect should matter more in the innovative sectors, we test:

Hypothesis 2: Stronger dismissal laws lead to relatively more innovation in the innovation-

intensive industries than in the traditional industries.

Since other aspects of labor laws do not have this ex-ante incentive effect, we also test:

Hypothesis 3: Laws governing dismissal of employees influence innovation more than other

dimensions of labor laws.

We first provide evidence supporting the hypotheses by exploiting changes in dismissal laws

at the country level. We use data on patents issued by the United States Patent and Trademark

Office (USPTO) to U.S. and foreign firms as well as citations to these patents as constructed by

Hall, Jaffe and Trajtenberg (2001). We measure innovation using the number of patents applied for

(and subsequently granted), the number of all subsequent citations to these patents, and, as our

model predicts more risk-taking subsequent to the passage of stronger dismissal laws, the standard

deviation of citations. As our primary explanatory variable, we employ an index of dismissal laws

developed by Deakin et al. (2007). They construct this index by analyzing in detail every legal

change pertaining to dismissal of employees in five countries — U.S., U.K., France, Germany, and

India — over the period 1970–2006. The index takes into account not just the formal or positive

law but also the self-regulatory mechanisms that play a functionally similar role to laws in certain

countries. While using the Deakin et al. index forces us to limit our empirical analysis to only

a limited set of countries, these countries account for 70% of the patents filed with the USPTO

during our sample period.4

We conduct our tests at two levels of aggregation: (i) country-level, where we only exploit

variation in innovation across time within a country; and (ii) industry-level, where we exploit

variation both across time and across different industries within a country. The “industry” level

classification we employ is very granular and corresponds to around 500 “patent classes” that the

USPTO defines. To test Hypothesis 1, we first examine the correlation between our innovation

proxies in a given country and year and dismissal laws in a given country in the previous year, as

well as two years prior. In estimating this correlation, we control for unobserved heterogeneity at

the country-level (through country fixed effects), secular time trends and macro effects (through

3The lawsuit alleges that Activision fired West and Zampella after they completed the video-game developmentand before they received the royalties for their work.

4Due to very limited patenting with the USPTO, we exclude India from our tests. However, all results remainrobust to the inclusion of India in the sample.

2

year dummies) as well as several country- and industry-level variables. The presence of the country

and year fixed effects enables us to estimate this correlation as a difference-in-difference, i.e., the

before-after difference in innovation in a country and year in which there was a change in dismissal

laws vis-a-vis the before-after difference in a country and year where there was no change in dismissal

laws. We find that more stringent dismissal laws in a particular year are positively correlated with

innovation in the following years.

As a specific source of endogeneity, changes in a country’s government (i.e., changes in its

political leanings) may confound our results, as could the correlation of dismissal law changes with

economic growth and/or business cycles. To directly control for these sources of endogeneity in the

difference-in-difference tests, we re-run our basic panel regressions after including: (i) a time-varying

proxy for the political leanings of a country’s government; (ii) the GDP growth rate to control for

the effect of economic growth; and (iii) country-specific periods of business cycle contractions. We

find that the effect of dismissal laws on innovation remains robust.

Despite controlling for an exhaustive set of observable variables that may influence innovation

and the passage of dismissal laws, we are careful not to ascribe a causal interpretation to the

above correlation since the possibility remains that unobserved factors accompanying law changes

may lead to the correlation. As the centerpiece of our identification strategy, we undertake triple-

difference tests where we absorb all variation at the country-year level through country x year

fixed effects and identify the effect within industries in a country. This identification strategy

is motivated by Hypothesis 2 above, which predicts that the effect of dismissal laws should be

disproportionately stronger in industries that exhibit a greater propensity to innovate than in

other industries. To conduct these tests, we proxy an industry’s propensity to innovate using the

National Science Foundation’s measure of the number of R&D scientists and engineers employed

per thousand employees in an industry. By interacting this proxy with the dismissal law index,

we find that the coefficient on this interaction term is significantly positive, which implies that

the effect of dismissal laws is more pronounced in industries that have a greater propensity to

innovate. These tests serve two important purposes. First, they highlight the channel for the

main effect – the industry’s propensity to innovate. Second, they ensure that neither changes

in a country’s government nor economic growth, country-specific business cycles, nor any other

country-level variable that correlates with dismissal law changes accounts for our findings.

Having controlled for all possible omitted variables at the country level, we then undertake

triple-difference tests that account for possible placebo effects at the country, industry level. In

our model, the hypothesized effect of dismissal laws on innovation stems from the increased effort

by a firm’s employees due to the reduced possibility of hold-up. Since individual inventors are not

employed by a firm, this predicted effect of dismissal laws should not manifest for innovation by

individual inventors. However, individual inventions may be affected by other, possibly omitted,

country and industry-level variables in a similar fashion as innovation by firms. Therefore, stand-

alone inventors provide a control group whose innovative output should not be affected by changes

in dismissal laws. Following this intuition, we employ innovation by firms minus innovation by

3

individual inventors as the dependent variable to net out any confounding effects driven by omitted

variables at the country, industry level. Reassuringly, our results continue to remain strong. Both

sets of triple-difference tests together provide convincing evidence of the causal effect of dismissal

laws on innovation.

In other tests, we confirm that the direction of causality runs from dismissal laws to innovation

rather than vice versa. In our final set of cross-country tests, we shed light on Hypothesis 3. Since

Deakin et al. generate indices to measure other dimensions of labor laws as well (for example,

laws governing industrial action, or employee representation), we study the effect of these other

dimensions of labor laws and find that dismissal laws are the only one that have a consistently

positive and significant effect on innovation.

Together, we conclude that stronger dismissal laws encourage innovation. The effect is eco-

nomically significant. Since we identify the intended effects using specific law changes, consider a

typical law change as an example. The U.K. increased the procedural hurdles relating to dismissal

of employees in 1987; this change corresponded to an increase of 0.0378 in the dismissal law index.

Using our coefficient estimates from the country-level tests, we find that this specific law change

increased annual number of patents, citations, and standard deviation of citations by 1.3%, 1.6%,

and 2.2% respectively.

To complement the country- and industry-level tests above, we conduct firm-level tests focusing

on one country alone by analyzing a federal dismissal law change in the U.S. in the form of the

passage of the WARN Act in 1989. In these tests, we exploit the discontinuity introduced by the

fact that WARN was applicable only to firms with 100 or more employees. Specifically, we examine

the before-after difference in patents and citations for firms with employees in the range [100,110]

– i.e., firms affected by WARN – vis-a-vis firms with employees in the range [90,99], which were

unaffected by WARN. We classify firms based on the number of employees in 1987 (i.e., two years

before the law change took effect) to avoid any endogeneity stemming from the classification itself.

First, by analyzing WARN Act notices to all firms in California, we find that the sample included

many innovative companies, which suggests that dismissal presents a credible threat to employees

in innovative firms as well. Second, we confirm that WARN indeed impacted the likelihood of

employee layoffs. Then, we undertake tests that compare before-after differences in innovation

between the treatment and control group of firms. We find that compared to those unaffected by

the passage of WARN, affected firms more patents post WARN, which are also more widely cited.

We show visually and econometrically that the discontinuous effect occurs at the cut-off of 100

employees; we find no effects at placebo cut-offs such as 50 or 150. Finally, using a difference-

in-difference specification, we report that our results obtain in the entire sample of firms as well,

not only in firms with employment in the range of (90,110) employees. These tests provide robust

evidence that is complementary to our cross-country tests since WARN was applicable selectively

to some firms but not others (based on their employment figures) while other federal laws that may

have been contemporaneous did not have such a discriminatory effect.

The cross-country tests together with the WARN-based results complement the findings in

4

Acharya, Baghai, and Subramanian (2012), who show that the staggered adoption of common-

law exceptions to the “employment-at-will” principle (so-called “Wrongful-Discharge Laws”) in

several U.S. states resulted in more innovation and entrepreneurship by U.S. firms. Apart from

the different setting (cross-country and U.S. federal law changes vis-a-vis U.S. state law changes),

this study differs from Acharya, Baghai, and Subramanian (2012) in other key ways. First, since

the cross-country setting provides variation stemming from passage of other labor laws as well,

we are able to confirm here that dismissal laws are salient in engendering positive incentives for

innovation, while other dimensions of labor laws do not have this salutary effect. Second, since our

cross-country tests exploit country-level changes in dismissal laws, these time-series tests provide

point estimates of the effect of changes in dismissal laws on innovation using experiments of greatest

relevance to country-level policies concerned with promoting innovation.

Our evidence of the beneficial effects of dismissal laws on innovation supports recent arguments

for enhancing employment protection in light of the recent financial crisis and the extraordinary rise

in the number of long-term unemployed (see Rajan, 2010, for instance). The last few recessions

have witnessed “jobless recoveries” and it has been argued that such joblessness – attributable

largely to aggregate risk rather than individual-specific risk – has resulted in weaker consumption

and investment growth out of recessions. Our results suggest that extension of unemployment

benefits when aggregate risk in the economy is high, again as suggested in Rajan (2010), can

embolden individuals to retrain themselves for newer jobs (a form of risk-taking), boost aggregate

consumption and demand, and in turn, corporate investment.

II Theoretical motivation

Setup. While innovative projects involve considerable risks stemming from exploration, the pay-

offs from successful innovation can be significant. If employment contracts are incomplete, as

highlighted by the theory on property rights (Hart, 1995), innovative firms may armtwist employ-

ees that contributed to a successful innovation to appropriate a larger share of the ex-post surplus;

this idea forms the essence of our theoretical model.

A recent high-profile court case filed against the video-game company Activision by its former

employees Jason West and Vince Zampella highlights this issue. According to the lawsuit, “(the

plaintiffs) created for Activision two videogame franchises, Call Of Duty and Modern Warfare, that

became the most successful in the company’s – indeed, the industry’s – history [...] In November

2009, [they] delivered to Activision Modern Warfare 2 – a video game that has already been

responsible for over $1 billion in sales and was recently hailed by Activision itself as the largest

launch of any entertainment product ever. Just weeks before Messrs. West and Zampella were to

receive the royalties for their hard work on Modern Warfare 2, Activision fired them in the hope

that by doing so, it could avoid paying them (the royalties) they had rightfully earned.”5

5See e.g. the complaint filed by West and Zampella in 2010 against Activision in the Su-perior Court of the State of California (County of Los Angeles), and the resulting trial doc-umented at http://ve3d.ign.com/articles/news/54192/Activision-Counter-Sues-Fired-Infinity-Ward-Founders-Suit-Scanned-Broken-Down-Transcribed.

5

Since incomplete contracts have been highlighted as one reason for the existence of employment

protection legislation (Bagchi, 2003), we develop a model where the employer cannot commit to

a contract that prohibits her from holding up the employee ex-post. Specifically, consider a firm

(F ), which chooses either of two projects that differ in their degree of innovation. We denote the

“routine” project by R and the “innovative” project by I. The firm has an employee (E ) who

works on the project chosen by the firm.

Timing and sequence of events. There are three cash flow dates, t = 0, 1, 2. At date 0, the

firm recruits an employee and chooses to invest in either the innovative or the routine project. The

projects require the same initial investment and generate cash-flows at date 2.

At date 1, the employee exerts effort ej ∈ [0, 1] , which affects the project outcome. We assume

the effort to be observable but not verifiable. The employee incurs a personal cost which we assume

to bee2j2 . If the employee chooses effort ej in project j then the project generates a successful

innovation with probability ej . At date 1.5, i.e., before the actual cash-flows accrue at date 2, all

agents learn whether the project chosen at date 0 produced an innovation or not. Since contracts

are incomplete, at date 1.5, E is exposed to the possibility of hold-up by F . Once the project

has generated a successful innovation, F could threaten to fire E and develop the innovation by

employing an alternative employee E′ who has limited bargaining power vis-a-vis F . We model the

bargaining process between E and F as the 50 : 50 Nash Bargaining solution with outside options.

To derive the outside options endogenously, we model the following extensive-form game between

E and F . After learning whether the innovation was successful or not, F decides whether to retain

E or fire him. If F decides to fire E, then dismissal laws limit F ’s ability to fire E. We capture this

facet by assuming that F can successfully fire E with probability µ, where µ captures the stringency

of the dismissal laws. The absence of legal firing restrictions nests as a special case where µ = 0.

At date 2, cash-flows are realized and allocated based on bargaining outcomes at date 1.5.

For project j, j ∈ {I,R} , the cash-flow equals αj if the project yields a successful innovation,

and βj ≤ αj if the project fails to generate an innovation, where αj is scaled to be less than one:

βj ≤ αj < 1. We assume that while the firm can replace E with a new employee E′ and produce, the

employee cannot produce without the firm. We further assume the labor market to be competitive

with employees earning their reservation utility in equilibrium, which we normalize to zero. Finally,

the common discount rate equals zero.

II.A Innovative vs. routine project

Routine projects face risks mainly due to uncertainty in market demand and competition. In

contrast, innovative projects entail additional risks associated with the process of exploration and

discovery. Therefore, the key difference we model is that the innovative project is riskier than the

routine one. We capture this difference in a simple manner by assuming that: βR = αR = R,

βI = 0, αI = A. Corresponding with the higher risk, we assume that the payoff from innovation is

higher than that from the routine project: A > R.

6

II.B Analysis

Since the payoff from the routine project is fixed (βR = αR = R), it is not affected by employee

effort. Therefore, we focus on the game that results if the firm chooses the innovative project I.

In Online Appendix A, we analyze the bargaining at date 1.5 and show that E’s expected payoff

given a successful innovation equals 0.5µA – the more stringent the dismissal laws, the greater the

payoff. Since the probability of a successful innovation equals eI , E’s net expected payoff U (eI) is

given by U (eI) = eI · 0.5µA − 0.5e2I . Therefore, the equilibrium level of effort, which is chosen by

E to maximize U (eI) , is given by:

eFBI = arg maxeI

[eI · 0.5µA− 0.5e2

I

]⇒ e∗I = 0.5µA (1)

Thus, the equilibrium level of effort increases with the stringency of the dismissal laws. To highlight

the effect of contractual incompleteness, consider the case when contracts are complete and F can

incentivize E to choose effort to maximize the total surplus generated from project I:

eFBI = arg maxeI

[AeI − 0.5e2

I

]⇒ eFBI = A (2)

When contracts are complete, the efficient level of effort: (i) does not depend on the stringency of

dismissal laws since hold-up is immaterial; and (ii) is lower than the level of effort when contracts

are complete: e∗I < eFBI . Intuitively, when contracts are incomplete, the employer cannot commit

to not hold up the employee after finding out that the innovation is successful. Since increased

innovative effort by the employee increases the likelihood of successful innovation, the likelihood of

hold-up by the employer decreases the employee’s effort in the innovative project.

Finally, since the labor market is competitive, employees earn their reservation wage in equilib-

rium. Therefore, as shown by the Lemma in the Online Appendix A, at date 0, the firm chooses

the project that maximizes the total surplus, which we denote by Wj , where jε{I,R}. While the

routine project’s joint payoff WR equals R, the joint payoff of the innovative project is given by:

W ∗I = Ae∗I − 0.5 (e∗I)2 = 0.125A2µ (4− µ) (3)

II.C Empirical predictions

Proposition 1 As dismissal laws become more stringent, the innovative effort exerted by an em-

ployee increases, which brings his effort closer to the first-best level. Formally,de∗Idµ >

deFBIdµ .

As dismissal laws become more stringent, the employee has a greater defense against hold-up

by the employer, which increases the outside option and thereby increases his share of the surplus

generated from a successful innovation. Therefore, when contracts are incomplete, more stringent

dismissal laws increase the employee’s innovative effort.

Proposition 2 As dismissal laws become more stringent, the employee’s effort in the innovative

7

project increases disproportionately more relative to the increase in the routine project. Formally,de∗Idµ >

de∗Rdµ .

Given our assumption that the routine project generates the same cash-flows whether the inno-

vation is successful or not, it follows that effort in the routine project equals a constant. Therefore,

we can infer from Proposition 1 that an increase in the stringency of dismissal laws disproportion-

ately increases the employee’s effort in the innovative project when compared to the increase in his

effort in the routine project.

Proposition 3 An increase in the stringency of dismissal laws increases the value of the innovative

project disproportionately more than the value of the routine project. Formally, dW∗I

dµ > dW∗R

dµ .

Proposition 4 Under the parametric restriction that the payoff from innovation is greater than a

threshold, there exists a µ ∈ (0, 1) such that the value from the routine project is higher than the

value from the innovative project when dismissal laws are not stringent (µ ≤ µ), and the reverse is

true when such laws are stringent (µ > µ). Formally, µ Q µ⇒W∗I (µ) QW

∗R (µ) .

The intuition for both propositions above follows directly from Proposition 2. When dismissal

laws are more stringent, the underinvestment in effort becomes disproportionately lower for the

innovative project than for the routine project. Therefore, the ex-ante expected surplus from

undertaking an innovative project increases disproportionately more than that from undertaking

a routine project. Thus, the increased employee effort in innovation generated by dismissal laws

translates into a positive effect on the expected firm value from innovation as well.

II.D Discussion

A key assumption we have made is that the firm and the employee cannot write a complete con-

tract. A natural question arises: Could parties commit to contractual features in the employment

contract, such as generous severance packages, to avoid inefficiencies stemming from contractual

incompleteness? Note that, as Hart (1995) explains, any ex-ante contractual features cannot lead

to credible commitment against hold-up when contracts are incomplete as in our setting. Empirical

evidence also indicates that for employees that do not constitute senior management in a firm, such

severance packages are quite uncommon.6 In innovative firms, this rarity of severance payments

below the level of senior management is consistent with the argument in Manso (2011), who shows

that even when complete contracts can be written, the firm may find it prohibitively costly ex ante

to commit to not fire its employees ex post.

Note that the assumption of 50:50 Nash Bargaining is not crucial to our results. Readers may

argue that giving E the full share of the surplus can alleviate the underinvestment problem that

is at the heart of our results. However, this is not the case for two reasons. First, since employees

6Narayanan and Sundaram (1998), who examined a combined sample of Fortune 1000 and S&P 500 companiesfrom the year 1995, find that while 55% of the firms had a “golden parachute” agreement with top management, only7% of the firms had “tin parachutes”, i.e., severance agreements for employees who are not officers of the company.Furthermore, such “tin parachutes” are limited to change-of-control events such as a merger or acquisition.

8

are risk averse and financially constrained, giving all the surplus to the employee may neither be

optimal (due to employees’ risk aversion) nor be feasible (due to employees’ financing constraints).

Second, though for simplicity we did not model the investment by the firm, the likelihood of a

successful innovation would be affected by the firm’s investment as well. If we allow for the firm’s

investment to also affect the likelihood of a successful innovation, giving all the surplus to the

employee would rob the firm of any incentives.

III Cross-country evidence

We first provide evidence of the effect of dismissal laws on innovation by exploiting changes in

dismissal laws at the country level.

III.A Data and proxies

Proxies for innovation: We employ the number of patents and citations to these patents as prox-

ies for innovation. While patents proxy successful innovation, the simple count of patents does not

distinguish breakthrough innovations from less significant or incremental technological discoveries.7

In contrast, citations capture the economic importance and drastic nature of innovation. Recall

that our theoretical argument implies that the passage of dismissal laws should lead to greater

employee effort in innovative projects, thereby enhancing the likelihood of successful innovation

that is value-enhancing to the firm. Since patent citations capture the economic importance of the

innovation, they enable us to proxy for the value-enhancing aspect of dismissal laws on innovation.

Since in our model the innovative project differs from the routine project with respect to project

risk, we also employ the standard deviation of citations to capture the riskiness of the innovative

project vis-a-vis the routine one.

We follow Acharya and Subramanian (2009) in using U.S. patents to proxy innovation by

international firms. To construct these proxies for innovation, we use data on patents filed with the

U.S. Patent Office (USPTO) and the citations to these patents, compiled in the NBER Patents File

(Hall, Jaffe and Trajtenberg, 2001). We exploit the technological dimension of the data generated

by “patent classes.” According to Kortum and Lerner (1999), the USPTO assigns patents to about

500 patent classes to facilitate future searches of the prior work in a specific area of technology.

We follow the practice in the patent literature in dating the patents by the year in which they

were applied for. This avoids anomalies that may be created due to the lag between the date of

application and the date of granting of the patent (Hall, Jaffe and Trajtenberg, 2001). Note that

although we use the application year as the relevant year for our analysis, the patents appear in

the database only after they are granted. Hence, we use the patents actually granted (rather than

the patent applications) for our analysis.

Dismissal law changes: We exploit the time-series variation generated by changes of dismissal

laws within countries. We use a comprehensive list of dismissal law changes from Deakin et al.

7Pakes and Shankerman (1984) show that the distribution of the importance of patents is extremely skewed, i.e.,most of the value is concentrated in a small number of patents. Hall, Jaffe and Trajtenberg (2005) among othersdemonstrate that patent citations are a good measure of the value of innovations.

9

(2007), who analyze in detail the evolution of employment protection legislation across five countries

from 1970 to 2006 and generate a labor law index.8

The Deakin et al. index offers several advantages. First, the long time-series, which captures

comprehensively all country level changes in dismissal laws, enables us to conduct tests that alleviate

econometric concerns that may otherwise be a problem in a cross-country setting. Second, their

categorization of labor laws into different components – dismissal laws being one of them – allows us

to assess the impact on innovation of dismissal laws vis-a-vis other categories of labor laws. Third,

Deakin et al. (2007) take into account not only formal laws but also self-regulatory mechanisms,

which makes their index particularly comprehensive with respect to the range of rules analyzed.9

Finally, the numerical values reported in their index are complemented by a detailed country-level

description of all the relevant law changes in each country. Though the Deakin et al. index is

available only for five countries – U.S., U.K., France, Germany and India – focusing on a limited

set of countries does not represent a substantial omission for the purposes of our analysis as these

five countries account for 70% of patents filed with the USPTO.10

To examine the effect of laws governing dismissal of employees on innovation, we focus on the

“Regulation of Dismissal” sub-index. We describe the individual components of this sub-index

(hereafter “the Dismissal Law index”or “Regulation of Dismissal index”) in Online Appendix B.

Figure 1 shows the evolution of the dismissal law index for the four countries in the sample; higher

values represent stricter laws governing dismissal. In the same graph, we plot the real GDP growth

rate for each country, as well as business cycle troughs. It is clear from the graph that while stricter

dismissal laws are more likely to be passed in periods of economic contractions, this relationship is

not strong (the correlation equals -0.18). Nonetheless, we control for real GDP growth in our tests.

Figure 1 (together with Table B-1 from Online Appendix B, which discusses each dismissal

law change during the time period 1970-2006) indicates that the numerous legal changes provide

substantial time-series variation, which we exploit in our statistical tests.

Summary statistics: Panel A of Table I reports the summary statistics for our sample. For each

of the four countries in our sample, this table lists the mean, median, standard deviation, minimum,

and maximum for the number of patents filed, citations received by these patents, the standard

deviation of citations, as well as the dismissal law index. The Deakin et al. index is available from

1970 to 2006 while the patent data ends in 2002, so we terminate our sample in 2002.

8The Botero et al. (2004) index presents an alternative to the Deakin et al. (2007) index that we use. AlthoughBotero et al. (2004)’s index is constructed for 85 countries, the index is available only for the year 1997. Therefore,it is not suitable to investigate the causal impact of labor laws on innovation, which necessitates controlling forobservable and unobservable time-varying heterogeneity. Another alternative is the EPL measure constructed byNicoletti and Scarpetta (2001) for a set of OECD countries for the years 1990-1998. However, this index neitheroffers the cross-sectional comprehensiveness of the index constructed by Botero et al. (2004), nor the full extentof the longitudinal advantages of the index developed by Deakin et al. (2007). Furthermore, the EPL index onlymeasures the aggregate stringency of a country’s labor laws, while in this study we are interested in one particulardimension of these laws, namely dismissal rules.

9For example, in certain legal systems, collective bargaining agreements – which do not constitute formal law –play a functionally similar role to formally enacted laws.

10Due to very limited patenting with the USPTO, we exclude India from our tests. However, all results remainrobust to the inclusion of India in the sample.

10

III.B Results

III.B.1 Fixed-effects panel regressions

Country-level tests: First, we run fixed-effects panel regressions of the innovation proxies on the

dismissal law index:

yct = βc + βt + β1 ∗DismissalLawsc,t−k + βXct + εct (4)

where yct is the natural logarithm of a measure of innovation from country c in year t. βc and βt

denote country and application year fixed effects, respectively. DismissalLawsc,t−k denotes the

kth lag of the dismissal law index for country c, measuring the stringency of dismissal laws. β1

measures the impact of dismissal laws on our innovation proxies. Xct is a set of control variables.

The country fixed effects control for time-invariant unobserved factors at the country level. The

application year fixed effects account for global technological shocks; further, they allow us to

control for the problem stemming from the truncation of citations, i.e., citations to patents applied

for in later years would on average be lower than citations to patents applied for in earlier years.

As explained by Imbens and Wooldridge (2009), in (4) , β1 estimates the “difference-in-difference”

in a generalized multiple treatment groups, multiple time periods setting. Figure 2 illustrates the

difference-in-difference for the change in laws governing dismissal in the U.S. in 1989 with Germany

as the control group. In this figure, we plot across time the ratio of realized number of patents in

a particular year to that in 1989 – the year of the U.S. dismissal law change. We find that while

the number of patents are relatively in sync for U.S. and Germany until 1989, post 1989, these

measures for the U.S. break ahead of those for Germany.

Table II, Panel A, shows the results of the test of equation (4) using the logarithm of the number

of patents, number of citations to patents, and the standard deviation of citations as the dependent

variables.11 Columns 1–3 and 4–6 employ respectively the first and second lags of the dismissal

law index, which enables us to estimate the impact of dismissal law changes on innovation one and

two years after the change respectively. In tests that we omit in the interest of brevity, we also find

similar effects on innovation three years after the change. Overall, the coefficient of dismissal laws

is positive and significant, which indicates that stronger dismissal laws are positively correlated

with innovation, as suggested by Hypothesis 1.

Industry-level tests: Next, we exploit variation in innovation across industries by measuring

our innovation proxies at the country and industry (patent class) level using the following OLS

regressions:

yict = tβj←i + tβc + βi + βt + β1 ∗DismissalLawsc,t−k + β ·Xict + εict (5)

where yict is the natural logarithm of a measure of innovation for the USPTO patent class i from

11To calculate the standard deviation of citations at the country and year level, we proceed as follows: for eachcountry and year, we first sum the number of citations each firm receives; we then calculate the standard deviationof these citations per country and year.

11

country c in year t.12 The patent class fixed effects βi control for average differences in technological

advances across the different industries as well as time-invariant differences in patenting and citation

practices across industries. The year fixed effects βt control for average trends in innovation across

all countries. tβj←i denotes a time trend for the industry j to which patent class i belongs;13 tβc

denotes a time trend for country c. Since other country or industry-level factors accompanying the

dismissal law changes could lead to country-specific as well as industry-specific time trends, these

tests enable us to isolate better the pure effect of dismissal law changes on innovation. The other

variables are as defined in equation (4). Since the explanatory variable varies at the country, year

level and our innovation proxies are measured at the patent class level, in these tests, we estimate

standard errors that are clustered at the country, year level.

We also account for the following variables in these tests: (i) creditor rights changes (Djankov

et al., 2007, index of creditor rights) and log of real GDP; (ii) bilateral trade using the imports and

exports of a country with the U.S. in each year in each 3-digit ISIC industry;14 (iii) industry-level

comparative advantage using the ratio of value added in a 3-digit ISIC industry in a particular year

to the total value added by that country in that year.15 The results of these tests are reported in

Panel B of Table II. Crucially, we find that the overall effect of dismissal laws stays positive and

significant for all three innovation proxies.

III.B.2 Addressing identification concerns

Despite the fixed effects and country and industry-specific time trends, we cannot provide a

causal interpretation to the above correlation since residual unobserved factors accompanying law

changes may lead to the above correlation. First, to cater to their political constituencies, more left-

leaning governments may be inclined to strengthen labor laws (see e.g., Botero et al., 2004; Deakin

et al., 2007). Leftist governments may also be more likely to invest in education and other public

services, which may have a positive impact on innovation in a country. Therefore, other factors

coinciding with changes in government may spoil identification. Second, dismissal law changes

may be also correlated with GDP growth/business cycles in a country.16 On the one hand, lower

economic growth/troughs in the business cycle may lead to an increase in the stringency of dismissal

12To calculate the standard deviation of citations at the country, patent class, and year level, we proceed as follows:for each country, patent class, and year, we first sum the number of citations each firm receives; we then calculatethe standard deviation of these citations per country, patent class, and year.

13Since there are about 500 patent classes in all, we create the interaction of the industry dummies with theapplication year at the patent category level, which encompasses several patent classes. There are six patent categories.

14The data are from Nicita and Olarreaga (2006). We match the patent classes to the 3-digit ISIC using a two-stepprocedure: first, the updated NBER patent dataset (patsic02.dta on Brownwyn Hall’s homepage) assigns each patentto a 2-digit SIC. We then employed the concordance from 2-digit SIC to 3-digit ISIC codes. Since every patent isalready assigned to a patent class in the original NBER patent dataset, this completes our match from the patentclass to the 3-digit ISIC code.

15The data for these measures come from the United Nations Industrial Development Organization (UNIDO)’sstatistics.

16For instance, Saint-Paul (2002) asserts that a higher economic growth rate reduces the political support fordismissal laws. However, since incumbent workers are most fearful of losing jobs during periods of slow economicgrowth, the political support for dismissal laws should be high in such periods. As empirical support for his thesis,Saint-Paul (2002) points out that in many European countries employment protection increased in the early 1970sand proved very difficult to reduce in the 1980s since this was a period of slow growth.

12

laws. On the other hand, innovation should foster economic growth, as suggested by the endogenous

growth theory (Aghion and Howitt, 1992). Thus, any effect of economic growth/business cycles on

dismissal laws could hinder the identification as well.

In Table III we directly address the concerns stemming from these sources of endogeneity. Panel

A focusses on the aggregate country-level sample corresponding to equation (4), while Panel B uses

the disaggregated sample corresponding to equation (5). First, we control for the effect of changes

in a country’s government using a time-varying proxy for the political leanings of a country’s

government. We use the variable Government from Armingeon et al. (2008), which captures the

balance of power between left and right-leaning parties in a given country’s parliament.17 This

variable takes on values from one to five, with one denoting a hegemony of right-wing (and center)

parties, and five denoting a hegemony of social-democratic and other left parties. Numerically,

the variable Government is positively correlated with the dismissal law index (the correlation is

0.49).18 We also include GDP growth and country-specific indicators for periods of business cycle

contractions.

Overall, there is some evidence that within a country, innovation is greater under more left-

leaning governments. However, while the coefficient of Government is positive throughout, it is not

statistically significant in most of the specifications. Similarly, we find our proxies for innovation to

be mostly positively correlated with GDP growth, and negatively correlated with times of business

cycle contractions, though these correlations are significant only in some specifications. Crucially,

however, we observe that the coefficient on the dismissal law index remains positive and significant

in all instances. Comparing the coefficients with and without controlling for these sources of

endogeneity shows that accounting for the possible endogeneity of dismissal law changes does not

materially affect the economic magnitude of the documented effect.

III.B.3 Triple-difference tests controlling for all country-level variables

The previous tests account for important sources of endogeneity. However, the concern re-

mains that some unobservable time-varying country-level omitted variables that are correlated

with changes in dismissal laws may confound our results. To address these endogeneity concerns,

we conduct a test where we include country x year fixed effects, which allow us to account for all

omitted variables for each country, year combination in our sample. The identification strategy is

motivated by Hypothesis 2 in our model, in which we argue that the effect of dismissal laws should

be disproportionately stronger in industries that exhibit a greater propensity to innovate than in

other industries.

We proxy propensity to innovate using the National Science Foundation’s measure of the number

of R&D scientists and engineers employed per thousand employees in a (manufacturing) industry

17Armingeon et al. (2008) construct a Comparative Political Data Set, which is a collection of annual political andinstitutional data for 23 democratic countries for the period of 1960 to 2006. Our variable Government is denoted“govparty” in Armingeon et al. (2008).

18Table B-2 in Online Appendix B reports in detail the years in which the political leanings of elected governmentschanged, as well as the years in which dismissal laws were altered. The table shows that more left-leaning governmentsindeed tend to pass stricter dismissal laws.

13

in the U.S.19 Since the U.S. remains the front-runner in innovation, this measure comes close

to the efficient level of innovative intensity for any industry. Furthermore, given technological

commonalities, an industry that is innovation intensive in the U.S. will be so in another country

as well, which enables us to proxy innovation intensity for a particular non-U.S. industry using the

U.S. measure as well.

In this test, we interact the dismissal law index with the innovation intensity of an industry:

yict = βc,t + tβj←i + βi + β1 · (DismissalLawsc,t−1 ∗ InnovationIntensityi,US) + βXict + εict (6)

Note that the interaction term (DismissalLawsc,t−1 ∗ InnovationIntensityi,US) varies at the level

of industry i in country c in application year t. Since our dependent variable, yict, exhibits equivalent

variability, the coefficient of interest β1 is well-identified and measures the relative effect of dismissal

laws across industries that vary in their innovation intensity. Country-year fixed effects (βc,t) allow

us to control for all observed and unobserved variables at the country and year level. The resulting

triple-difference tests exploit the variation within industries depending on their innovation intensity.

Hypothesis 2 predicts that β1 > 0.

The results of this test are reported in Columns 1-3 of Table IV. The coefficient of the interaction

term β1 is positive and statistically and economically significant, indicating that the positive impact

of dismissal laws on innovation is more pronounced in innovation intensive industries. In Columns

4-6, we exclude U.S. from the sample and find that our results remain significant.

III.B.4 Triple-difference tests controlling for industry-level placebo effects

Next, we alleviate endogeneity concerns stemming from time-varying omitted variables at the

country- and industry-level by identifying a control group of innovating entities that would be

affected by such omitted variables but should be unaffected by dismissal law changes. In our

model, the hypothesized effect of dismissal laws on innovation stems from the increased dismissal

protection for firm employees. Dismissal law changes should not have an impact on individual

inventors, who are not employed by a firm. Therefore, they provide a relevant control group to

net out possible placebo effects. Based on this intuition, we conduct the following triple-difference

test, in which we examine the effect of dismissal laws on innovation by firms minus the innovation

generated by stand-alone inventors:

yict, firms− yict, individuals = tβj←i + tβc + βi + βc + βt + β1 ∗DismissalLawsc,t−1 + βXict + εict (7)

19The data for the innovation intensity measure are taken from Table A-54 of the 1993 National Science Founda-tion/SRS, Survey of Industrial Research and Development. For each of the 2-digit SIC manufacturing industries, wecalculate the average fraction of scientists employed over the 1983–1993 period. To merge the SIC industries to patentclasses, we use the assignment of SIC codes for each patent from the NBER patent file. Specifically, for all countriesavailable in the NBER patent file, we determine for each patent class the SIC that most patents were assigned toover the 1970–2002 period; that SIC is used as the representative SIC for that patent class. InnovationIntensityi,US

is available for 15 2-digit SIC manufacturing industries, or 245 patent classes in our sample.

14

where yict, firms and yict, individuals represent innovation by firms and individuals in a patent class i,

country c, and year t.20 Xict is the set of control variables, and tβj←i and tβc denote trends at the

industry- and country-level respectively.

In Columns 7 & 8 of Table IV, we find the coefficient β1 to be positive and statistically as

well as economically significant. These triple-difference tests enable us to control for any omitted

country- or industry-level variable that affects the passage of dismissal laws and affects innovation

performed by all agents in the economy.

In sum, we can conclude with a reasonable degree of certainty that country-level changes in

dismissal laws did indeed foster innovation by firms in that country and that our results are not

affected by possible endogeneity stemming from other country- or industry-level confounding factors

that may have coincided with the dismissal law changes.

III.B.5 Causality or reverse causality?

To examine possible reverse causality, we use the dismissal law change in the U.S. in 1989;

unlike in other countries, this change occurred at one isolated point in time, which makes it ideal to

address such a concern. We follow Bertrand and Mullainathan (2003) and decompose the change

in U.S. dismissal laws into three separate time periods: (i) Dismissal Law Change (-2,0), which

captures any effects from two years before to the year of the change; (ii) Dismissal Law Change

(1,2), which captures the effects in the year after the change and two years after the change; and

(iii) Dismissal Law Change (≥3), which captures the effect three years after the change and beyond.

Table V shows the results of these regressions. A positive and significant coefficient on Dismissal

Law Change (-2,0) could be symptomatic of reverse causation. While the coefficient in Column 3 is

small but positive and precisely estimated, we find that the coefficient is negative and statistically

significant in Column 1, and it is statistically insignificant in Column 2. In contrast, the coefficients

of Dismissal Law Change (1,2) and Dismissal Law Change (≥3) in Columns 1–3 show that while

the dismissal law change has a positive effect on the innovation proxies in the first two years, the

effect of the law change lasts three years and beyond. In fact, consistent with the long gestation

periods involved with innovative projects, we find the “long-run” effect to be larger than the effect

in the first two years.

III.B.6 Effect of other dimensions of labor laws

Next, we test our Hypothesis 3 that labor laws that affect the ex-post likelihood of an employee

being dismissed from employment matter more for innovation than other categories of labor laws.

For this purpose, we disaggregate the Deakin et al. (2007) labor law index into its components:

(i) Alternative employment contracts; (ii) Regulation of working time; (iii) Regulation of dismissal

– our “dismissal law index”; (iv) Employee representation; and (v) Industrial action.21 Table VI

presents results of these tests; the only dimension of labor laws which has a consistently positive

20Note that as individual-specific identifiers are not available in the patent data set (as opposed to firm-specificidentifiers), we cannot construct a measure for the standard deviation of citations for individual inventors.

21Note that while the correlation between different labor law components is positive and significant, the tests donot encounter any multi-collinearity problem.

15

and significant impact on innovation is the “regulation of dismissal” component.

III.B.7 Economic magnitudes

Having convinced ourselves of the causal effect of dismissal laws and innovation, we estimate

its economic magnitude. Since we identify the effect of dismissal laws using specific law changes,

consider, for example, the effect of the law change relating to procedural constraints on dismissal

in the U.K. in 1987 when it became harder for employers to avoid a finding of unjust dismissal in

case of a lack of due process. This law change corresponds to an increase of 0.0378 in the dismissal

law index. Using Columns 1-3 in Panel A of Table II, this law change corresponds to an increase in

annual number of patents, citations, and standard deviation of citations by 1.3%, 1.6%, and 2.2%

respectively.

IV Within-country evidence using the WARN Act

In the previous sections, we provided evidence of the impact of labor laws on innovation in a

cross-country setting. In this section, we present tests of our main hypothesis based on U.S. data

alone. These tests exploit a discontinuity introduced by the passage of the federal U.S. Worker

Adjustment and Retraining Notification (WARN) Act and do not rely on labor law index data.

Our tests in this section are aimed at removing any concerns about: (i) potential measurement

error arising from the use of U.S. patents to proxy innovation by international firms; and (ii)

measurement error stemming from the use of law indices.

IV.A An overview of the WARN Act

The WARN Act is a federal law that was enacted by the U.S. Congress on August 4, 1988, and

became effective on February 4, 1989.22 The WARN Act requires employers to give written notice

60 days before the date of a mass layoff or plant closing to: (i) affected workers; (ii) chief elected

official of the local government where the employment site is located; and (iii) the State Rapid

Response Dislocated Worker Unit. Subject to the law are private employers with 100 or more full-

time employees, or with 100 or more employees who work at least a combined 4,000 hours a week.23

In the case of non-compliance, employees, their representatives, and units of local government can

bring lawsuits against employers. Employers who violate the WARN Act are liable for damages in

the form of back pay and benefits to affected employees.

To investigate whether innovative companies were affected by the WARN Act, we obtained

WARN Act notices received by the Employment Development Department in California in 2009.

These included several innovative companies, which illustrates the fact that dismissal presents

22The details on the WARN Act reported in this section are drawn from the following two sources, un-less otherwise noted: United States Department of Labor – Employment & Training Administration (http ://www.doleta.gov/layoff/warn.cfm); and Levine (2007).

23Only layoffs classified as “mass layoffs” or “plant closings,” or layoffs of 500 or more full-time workers at asingle site of employment, are covered. A “plant closing” is defined as a closure of a facility within a single site ofemployment involving layoffs of at least 50 full-time workers. In the case of a “mass layoff,” an employer lays offeither between 50 and 499 full-time workers at a single site of employment, or 33% of the number of full-time workersat a single site of employment. For further details, see Levine (2007).

16

a distinct threat to employees in innovative firms as well.24 After discussing our data and test

methodology, we will provide additional evidence of the importance of the WARN Act by showing

the impact of its passage on employment fluctuations.

The requirement of prior notification to local government together with penalties for non-

compliance imply that the WARN passage increases the hurdles faced by employers when dismissing

employees. This effect is in line with the effect of dismissal laws as discussed in our theoretical

motivation. Therefore, we expect WARN to have the predicted positive effect on innovation.

IV.B Data and sample

In order to examine the effect of the passage of the WARN Act on innovation, we match the

NBER patents file to Compustat data.25 As before, to construct proxies for innovation, we employ

patents filed with the USPTO and citations to these patents, compiled in the NBER Patents File

(Hall, Jaffe and Trajtenberg, 2001). Since it is not possible to construct time varying measures of

the standard deviation of citations at the firm level, we exclude this dependent variable from the

WARN analysis. The summary statistics for the main variables used in these firm-level tests are

displayed in Panel B of Table I.

IV.C Difference-in-difference tests

In these U.S. based tests, we exploit the discontinuity introduced by the WARN Act, which

was applicable only to firms with 100 or more employees. Our identification strategy is based on

comparing U.S. firms that were affected by the law change (firms with 100 or more employees) to

U.S. firms that were not (firms with less than 100 employees). Figure 3 illustrates this discontinuity

in innovation caused by the passage of the WARN Act. Here, we plot the before-after difference in

the number of patents filed by each firm against the number of employees in the firm in 1987, i.e.

one year before the passage of the WARN Act. The figure also shows linear fits over the [50,99]

and [100,150] employee ranges as well as a local polynomial smoothing fit over the same range. The

figure illustrates clearly that the break in innovation occurs at the threshold of 100 employees.

We first investigate the effect for the entire sample of firms using the following regression:

yit = βi + βt + β1 ∗ (Over100)i,1987 ∗ (After1988)t + β ·Xit + εit (8)

where yit is a proxy for innovation by firm i in year t. (Over100)i,1987 is a dummy taking the value

of one if a firm has ≥ 100 employees in the year 1987, i.e. two years before the passage of the WARN

24This list includes Adobe Systems Incorporated, AT&T company, Comcast Cable, FOX Interactive Media, Inc.,Genentech, Inc., Henkel Corporation, National Semiconductor Corporation, NEC Electronics America, Inc., Palm,Inc.; SAP America, Inc., Seagate Technology LLC, Siemens, Stanford University, Sun Microsystems, Inc., Symantec,The Boeing Company, Valeant Pharmaceuticals International, Virgin Mobile USA, Yahoo! Inc., and many others.Source: http : //www.edd.ca.gov/Jobs and Training/warn/eddwarnlwia09.pdf

25Each assignee in the NBER dataset is given a unique and time-invariant identifier. We match the U.S. assigneenames in the NBER patent dataset to the names of divisions and subsidiaries belonging to a corporate family fromthe Directory of Corporate Affiliations. We then match the name of the corporate parent to Compustat. Additionally,we augment our match of the U.S. assignee names to the Compustat parent with the recent gvkey-assignee matchdeveloped by NBER. See https://sites.google.com/site/patentdataproject/Home/downloads for the details about thisnew match.

17

Act, and 0 otherwise. By using employment information from the year 1987 only, we avoid the

endogeneity stemming from group classification due to the layoffs. This is a useful instrument for

two reasons. First, it is unlikely that the WARN Act had an impact on employment two years prior

to its passage. Second, whether a firm had more than 100 employees in 1987 is a good predictor for

the other years. (After1988)t is a dummy taking the value of one after the passage of the WARN

Act (i.e. for the years 1989-1994). The firm dummies βi control for any residual time-invariant

heterogeneity among firms. The year dummies βt account for general macro-economic factors. Xit

represents the set of control variables which include Size and Market-to-Book ratio.26 The sample

covers twelve years around the passage of the WARN Act (from 1983-1994). In all the regressions,

we cluster standard errors at the firm level.

Table VII shows the results. In Columns 1–4, we employ logs of the number of patents and

citations respectively as the dependent variables. While Columns 1 & 2 do not include additional

variables, we control for firm size and the market-to-book ratio in Columns 3 & 4. In line with

Hypothesis 1, we find that the WARN Act had a positive and significant impact on U.S. firm-level

innovation. Based on estimates from Column 1, compared to the control group, annual patents

increased by 18.4% for the treatment group of firms, with an even larger effect for citations.

According to our model, the positive effect of dismissal laws on innovation results from the

positive effect that these laws have on employee effort. Unlike our cross-country set-up, our sample

here is constructed at the firm level. Therefore, we can investigate whether the passage of WARN

had an effect on employee effort in innovative projects. For this purpose, we normalize our proxies

for innovation using the number of employees in a firm. In Columns 5-8 of Table VII, we report

the results using ln(patents/employees) and ln(citations/employees) as the dependent variables.

Here, we find that both patents and citations per employee increase after the passage of WARN

for the “treatment” group of firms. This finding shows that the theoretical backdrop finds support

not only in the result linking dismissal laws to innovation, but also in the specific mechanism we

conjecture to be at play, i.e. that the positive effect of dismissal laws on innovation results from

the positive effect that these laws have on employee effort.

IV.D Regression-Discontinuity Tests

To fully exploit the discontinuity due to the WARN Act and thereby provide the cleanest

evidence of our hypotheses, we focus on firms in the range [90,110].27 To ascertain that results

are not spurious, as placebo tests, we also test for any effects on innovation by using cutoffs of

50 and 150 employees and a sample of firms with employees in the range [40, 60] and [140, 160]

respectively. We proceed by first showing that WARN indeed had a dampening effect on employee

dismissals in affected firms. We then estimate the effect of WARN on innovation.

26Market-to-Book is the market value of assets to total book assets. Market value of assets is total assets (Compustatdata item at) plus market value of equity minus book value of equity. The market value of equity is calculated ascommon shares outstanding (csho) times fiscal-year closing price (prccf). Book value of equity is defined as commonequity (ceq) plus balance sheet deferred taxes (txdb). Size is the natural logarithm of sales (sale).

27Since the number of firms in the [99,101] range is quite limited, we employ the expanded window [90,110].

18

IV.D.1 Test design: WARN Act and employee layoffs

When examining the effect of the WARN Act on innovation, a key question that arises is whether

the WARN Act indeed imposed a binding constraint for innovative firms. The anecdotal evidence on

employee layoffs discussed above provided preliminary evidence of the same. To test this formally,

we define employee layoffs to have occurred in firm i in year t if the number of employees in that

year is lower than that in the previous year. We then estimate the following linear probability

model for the twelve years surrounding the passage of the WARN Act (1983-1994):

Ind(Empi,t−Empi,t−1 < 0) = βt+β1 ·(Over100)i,1987∗(After1988)t+β2 ·(Over100)i,1987+εit (9)

where Ind(Empi,t − Empi,t−1 < 0) is a binary variable taking on a value of one in case of a net

employment reduction in firm i from year t − 1 to year t. The other variables are as defined in

equation (8). Since employee layoffs due to the WARN Act do not exhibit much within-firm vari-

ation, we do not include firm-fixed effects. However, to control for average differences in employee

layoffs across years, we include the year fixed effects βt.

Column 1 in Panel A of Table VIII reports the results of the tests of equation (9) for firms

having employees in the range [90,110] in 1987. We find that the passage of WARN decreased the

likelihood of layoffs in the affected firms. Compared to the control set of firms in the range [90,99],

the before-after difference in the likelihood of employee layoffs decreased by 25% for the treated

firms in the range [100,110]. Columns 2-5 in Panel A of Table VIII show the results for the effect

of WARN on innovation. First, in Columns 2-3 of Panel A, we report the results of tests for our

proxies of aggregate innovation using the log of the number of patents and citations respectively

as the dependent variables. In line with Hypothesis 1, we find that the strengthening of dismissal

laws via the WARN Act had a positive and significant impact on U.S. firm-level innovation.

To assess the economic magnitude of the effect of the WARN Act, note that the median firm

with employees in the range [90, 110] files one patent per year in our sample; the median firm in

this range also receives 17 cumulative citations. After passage of the WARN Act, affected firms

file about one additional patent every two years after the passage of WARN. Furthermore, these

firms receive 14 additional citations in all. Since citations are a stock measure, the 14 additional

citations post WARN translate into two additional citations per year on average after the passage

of the WARN Act.

We test Hypothesis 2 in Columns 4-5 of Panel A by using ln(patents/employees) and ln(citations/employees)

as the dependent variables. Here as well, we find that both patents and citations per employee

increase after the passage of WARN for the “treatment” group of firms; however, the increase is

only statistically significant for the citations-based measure.

Panels B and C of Table VIII show the results for the placebo tests using only firms with

employment in the range [40, 60] and [140, 160] respectively in 1987. In each of these panels,

Column 1 shows the effect on employee layoffs while Columns 2-3 show the results for the log of

the number of patents and citations respectively, and Columns 4-5 report the results using log of

19

the number of patents and citations per employee respectively. In both these panels, we can infer

that there was no differential effect at the corresponding placebo cutoffs. This provides reassurance

that the positive effect of WARN on innovation documented in Panel A is not spurious.

IV.E Discussion

Apart from not suffering from concerns relating to unobserved factors at the country level, the

above tests based on WARN offer other attractive advantages. First, since our sample for the

WARN tests ended in 1994, they enable us to conclude that our results on the positive effect of

dismissal laws on innovation are not driven by any spurious effects that patent reforms motivated

by the General Agreement on Tariffs and Trade (GATT) may have had.28

Second, the tests based on the WARN Act mitigate effects of any other contemporaneous factors

that may confound our results. This strength of the WARN based tests stems from a combination

of three factors. First and foremost, since the firms are separated into treatment and control

groups based on the number of employees, any unobserved factor that affects all firms uniformly

(i.e. irrespective of employment figures) cannot be driving our results. Nevertheless, as a second

line of defense, we have used firm-fixed effects to account for time-invariant effects of unobserved

factors, in general, and firm size, in particular. Second, we have performed regression-discontinuity

tests to focus on firms just above and below the employment cut-off relevant for WARN. Third,

in the difference-in-difference tests, we have included firm size to account for any time-varying

correlation of any unobserved factors with firm size. Therefore, laws or policy changes or any other

unobserved factors that may influence innovation cannot affect the results unless they resemble

WARN in discriminating based on the size of the workforce.

Related to the above, the WARN tests also alleviate concerns that our results may be affected

by the coinciding of the post WARN period with the recession in the early 1990s. To the extent

that this recession slowed down the average pace of innovation, the application year fixed effects

should capture this effect. Finally, since firms of similar sizes should have felt the effect of the

recession similarly, the regression-discontinuity specification provides confirmation that our results

are not affected by the recession in the 1990s.

Finally, the WARN Act was not intended to specifically encourage innovation or economic

growth.29 Therefore, our tests above can reasonably be interpreted as a truly causal effect of the

WARN Act passage on innovation.

28Under the GATT changes, an unexpired issued patent or a patent application pending on June 8, 1995, has aterm of protection that is the longer of 17 years from the date of issuance of the patent or 20 years from the filingdate of the patent application. For applications filed on or after June 8, 1995, the patent life is now twenty years,measured from the earliest patent application. However, since our sample for the WARN tests ended in 1994, ourresults are not driven by GATT related changes.

29Brugemann (2007) examines various articles in the business press that document the events preceding and fol-lowing the WARN Act. He does not find any evidence arguing that the Act was aimed at improving a specific aspectof the U.S. economy.

20

V Related literature

Our study complements the findings in Acharya, Baghai, and Subramanian (2012), who exploit

state-level variation in dismissal laws within the U.S. to demonstrate their positive effects on in-

novation. They show that the staggered adoption of common-law exceptions to the “employment-

at-will” principle (so-called “Wrongful-Discharge Laws”) in several U.S. states resulted in more

innovation by U.S. firms.

Our paper also contributes to the body of literature that examines the effect of laws governing

the employer-employee relationship (Botero et al., 2004, Atanassov and Kim, 2007, Besley and

Burgess, 2004, Bassanini, Nunziata and Venn, 2009). In contrast to these studies which document

negative effects of labor laws, our study finds that some types of stringent labor laws can motivate

a firm and its employees to pursue value-enhancing innovative activities. Our study resembles

that of Menezes-Filho and Van Reenen (2003) in documenting the positive effect of labor laws.

However, while Menezes-Filho and Van Reenen (2003) focus on laws governing unions, we examine

all dimensions of labor laws and pay particular attention to laws governing dismissal of employees.

Our study relates to that of MacLeod and Nakavachara (2007), who develop a theoretical

model and provide empirical evidence that the passage of wrongful discharge laws across several

U.S. states enhances (reduces) employment in industries requiring high (low) relationship specific-

investment. Garmaise (2007) uses legal enforcement of employee non-compete agreements across

U.S. states as a proxy for laws that limit human capital mobility and finds that such laws enhance

executive stability. However, in contrast to our results, such limits on human capital mobility

reduce firm-level investments in Research and Development in his study. One way to rationalize

our findings with his is that non-compete agreements induce employee stability by restricting their

freedom from departure when they wish to leave firms; in contrast, dismissal laws induce employee

stability by restricting the ability of firms to fire employees. Hence, in our setting, lower mobility

encourages innovative pursuits since employees are less likely to be fired in case innovative projects

fail (including due to sheer bad luck), thereby making innovation more profitable from an ex ante

standpoint for firms too. Saint-Paul (2002) argues theoretically that employment protection may

alter the pattern of specialization and innovation in favor of mature goods. Finally, Lerner and

Wulf (2007) examine U.S. publicly listed firms with centralized R&D units and find that long-term

incentives provided to corporate R&D heads are associated with greater firm-level innovation.

VI Conclusion

In this paper, we showed that firm-level innovation is causally determined by laws governing

the ease with which firms can dismiss their employees. Since the outcomes of innovation are un-

predictable, they are difficult to contract ex ante (Aghion and Tirole, 1994), which renders private

contracts to motivate innovation susceptible to renegotiation. Such possibility of renegotiating con-

tracts dilutes their ex ante incentive effects. Since laws are considerably more difficult for private

parties to alter than firm-level contracts, legal protection of employees in the form of stringent

dismissal laws can introduce the time-consistency in firm behavior absent with only private con-

21

tracts. Specifically, dismissal laws limit employers’ ability to hold up innovating employees after

an innovation is successful. By reducing the possibility of hold-up, these laws enhance employees’

innovative efforts and encourage firms to invest in risky, but potentially mould-breaking, projects.

After formalizing this intuition, we provided supporting empirical evidence using patents and cita-

tions as proxies for innovation and dismissal law changes across countries. We complemented the

cross-country results with a regression discontinuity design examining firm-level innovation after

the passage of the 1989 WARN Act in the U.S.

Assessing the aggregate welfare implications of labor laws is an important topic for future

research. Our study highlights one important positive effect of dismissal laws, namely their ability

to spur innovation and economic growth, that must be factored into such an assessment. Since

endogenous growth theory (Aghion and Howitt, 1992) posits that firm-level innovation fosters

country-level economic growth, examining the effect of labor laws on economic growth through

innovation and entrepreneurship represents one important step in understanding the net effect of

labor laws. We suggest these as areas for future research.

References

[1] Acharya, V., R. Baghai and K. Subramanian, 2012, “Wrongful Discharge Laws and Innova-tion,” Working paper, New York University Stern School of Business.

[2] Acharya, V., K. Subramanian, 2009, “Bankruptcy Codes and Innovation,” Review of FinancialStudies, 22(12), 4949–4988.

[3] Aghion, P. and P. Howitt, 1992, “A Model of Growth Through Creative Destruction,” Econo-metrica, 60(2), 323–352.

[4] Aghion, P. and J. Tirole, 1994, “The Management of Innovation,” The Quarterly Journal ofEconomics, 109(4), 1185–1209.

[5] Armingeon, K., M. Gerber, P. Leimgruber, and M. Beyeler, 2008, “Comparative Political DataSet 1960–2006, ” Institute of Political Science, University of Berne.

[6] Atanassov, J., and E. H. Kim, 2007, “Labor Laws and Corporate Governance: InternationalEvidence from Restructuring Decisions,” Working Paper, Ross School of Business.

[7] Bagchi, A., 2003, “Unions and the Duty of Good Faith in Employment Contracts,” Yale LawJournal 112, 1881–1910.

[8] Bassanini, A., L. Nunziata and D. Venn, 2009, “Job protection legislation and productivitygrowth in OECD countries, ” Economic Policy 24(58), 351–402.

[9] Bertrand, M. and S. Mullainathan, 2003, “Enjoying the Quiet Life? Corporate Governanceand Managerial Preferences,” Journal of Political Economy, 111, 1043–1075.

[10] Besley, T. and R. Burgess, 2004, “Can Labor Regulation Hinder Economic Performance?Evidence from India,” Quarterly Journal of Economics, 119(1), 91–134.

[11] Botero, J., Djankov, S., La Porta, R., F. Lopez-De-Silanes and A. Shleifer, 2004, “The Regu-lation of Labor,” Quarterly Journal of Economics, 119(4), 1339–1382.

[12] Brugemann, B., 2007, “Employment Protection: Tough to Scrap or Tough to Get?,” EconomicJournal, 117, 386–415.

[13] Deakin, S., P. Lele, and M. Siems, 2007, “The evolution of labour law: Calibrating andcomparing regulatory regimes,” International Labour Review 146(3-4), 133–162.

[14] Djankov, S., C. McLiesh and A. Shleifer, 2007, “Private Credit in 129 Countries,” Journal ofFinancial Economics, 85(2), 299–329.

22

[15] Ederer, F. and Manso, G. 2010. “Is Pay-for-Performance Detrimental to Innovation?,” Paperpresented at the Entrepreneurship Finance and Innovation Conference jointly organized bythe Kauffman Foundation and the Review of Financial Studies.

[16] Garmaise, M., 2007, “Ties that Truly Bind: Non-competition Agreements, Executive Com-pensation and Firm Investment,” UCLA Anderson working paper.

[17] Hall, B. H., A. Jaffe and M. Trajtenberg. 2001. “The NBER Patent Citations Data File:Lessons, Insights and Methodological Tools,” NBER working paper.

[18] Hall, B., A. Jaffe, and M. Trajtenberg, 2005, “Market value and patent citations,” RANDJournal of Economics 32, 101–128.

[19] Hart, Oliver. 1995. Firms, Contracts, and Financial Structure (Oxford: Clarendon Press).[20] Imbens, Guido M. and Jeffrey M. Wooldridge, 2009, “Recent Developments in the Economet-

rics of Program Evaluation,” Journal of Economic Literature, 47(1), 5–86.[21] Kortum, S. and J. Lerner, 1999, “What is behind the recent surge in patenting?,” Research

Policy, 28, 1-22.[22] Lerner, Josh, and Julie Wulf, 2007, “Innovation and Incentives: Evidence from Corporate

R&D,” The Review of Economics and Statistics, 89(4), 634–644.[23] Levine, L., 2007, “The Worker Adjustment and Retraining Notification Act (WARN),” Work-

ing Paper, Congressional Research Service.[24] MacLeod W.B. and V. Nakavachara, 2007, “Can Wrongful Discharge Law Enhance Employ-

ment?,” The Economic Journal, 117, 218—278.[25] Manso, Gustavo, 2011, “Motivating Innovation,” Forthcoming, Journal of Finance.[26] McKinsey Global Institute, 1997, France and Germany (Washington, DC: McKinsey).[27] Menezes-Filho, N., and J. Van Reenen, 2003, “Unions and Innovation: A Survey of the The-

ory and Empirical Evidence,” in John T. Addison and Claus Schnabel, eds., InternationalHandbook of Trade Unions (Edward Elgar Publishing Ltd), 293–334.

[28] Narayanan, M., and A. Sundaram, 1998, “A Safe Landing? Golden Parachutes and CorporateBehavior,” Working Paper, University of Michigan.

[29] Nicita A. and M. Olarreaga, 2006, “Trade, Production and Protection: 1976-2004,” WorldBank Economic Review, 21(1).

[30] Nicoletti, G. and S. Scarpetta, 2001, “Interaction between Product and Labour Market Regu-lations: Do They Affect Employment? Evidence from the OECD Countries,” Conference on“Labour Market Institutions and Economic Outcomes.” Lisbon: Bank of Portugal, June 3–4,2001.

[31] Pakes, A., and M. Shankerman, 1984, “The rate of obsolescence of patents, research gestationlags, and the private rate of return to research resources,” in Zvi Griliches, ed., R&D, Patentsand Productivity (University of Chicago Press), 98–112.

[32] Rajan, R., 2010, Fault Lines: How Hidden Fractures Still Threaten the World Economy(Princeton University Press).

[33] Saint-Paul, Gilles, 2002, “Employment protection, international specialization, and innova-tion,” European Economic Review, 46, 375–395.

[34] Tian, X., and T.Y. Wang, 2011, “Tolerance for Failure and Corporate Innovation,” Forthcom-ing at the Review of Financial Studies.

23

Fig

ure

1:

Dis

mis

sal

Law

s,G

DP

Gro

wth

,an

dB

usi

ness

Cycle

Tro

ugh

s.

The

figure

show

sth

est

rength

of

the

“R

egula

tion

of

Dis

mis

sal”

(solid

line)

plo

tted

again

stth

era

teof

real

GD

Pgro

wth

(dash

edline)

for

agiv

enco

untr

yand

yea

r.T

he

ver

tica

ldott

edlines

den

ote

busi

nes

scy

cle

troughs.

Hig

her

valu

esof

the

dis

mis

sal

law

index

indic

ate

more

emplo

ym

ent

pro

tect

ion,

i.e.

,st

rict

erla

ws.

The

dis

mis

sal

law

index

data

isfr

om

Dea

kin

etal.

(2007).

Data

on

GD

Pgro

wth

isfr

om

the

Pen

nW

orl

dT

able

.C

ountr

y-s

pec

ific

busi

nes

scy

cle

data

isfr

om

the

Eco

nom

icC

ycl

eR

esea

rch

Inst

itute

(EC

RI)

.

24

Figure 2: U.S. WARN Act and Innovation: U.S. vs Germany.

This figure shows a plot across time of the ratio of the realized number of patents in a particular year tothat in 1989, the year the U.S. WARN Act became effective. The continuous line shows the ratio for theU.S. while the discontinuous line shows the same for Germany, which experienced no dismissal law changein the time interval examined. The index data is from Deakin et al. (2007).

Figure 3: WARN Act and Innovation by US Firms.

This figure shows the effect of the passage of the WARN Act on innovation by treatment group (firms with≥ 100 employees) and control group (firms with < 100 employees). Specifically, we plot the before-afterdifference (with respect to the WARN passage in 1989; the sample spans the years 1983–1994) in patents foreach firm against its number of employees in 1987. The solid curves indicate a polynomial smoothing fit ofthis before-after difference in patenting for treated and control firms, respectively; the dashed lines show alinear fit for the same data.

25

Table I: Summary Statistics.

Panel A (Cross-Country Sample) of the table gives summary statistics for the following variables per country: number ofpatents, number of citations, standard deviation of citations, and the dismissal law index. The innovation measures are definedat the country, patent class, and year level, while the dismissal index varies at the country and year level. The data span theyears 1970–2002.Panel B (U.S. WARN Sample) of the table gives summary statistics for the main variables used in the single-country U.S.WARN tests. Market-to-Book ratio is the market value of assets to total book assets. Market value of assets is total assetsplus market value of equity minus book value of equity. The market value of equity is calculated as common shares outstandingtimes fiscal-year closing price. Book value of equity is defined as common equity plus balance sheet deferred taxes. Size is thenatural logarithm of sales. The sample spans 1983–1994.Patent data is from the NBER Patents File (Hall, Jaffe and Trajtenberg, 2001). The labor law index data is from Deakin etal. (2007). Firm-level data is from Compustat.

Panel A: Cross-Country Sample

United States

Obns. Mean Median Std. Devn. Min. Max.Number of patents 13079 91.865 47 145.452 1 2879Number of citations 13079 653.759 267 1114.824 0 12116Standard deviation of citations 12611 18.748 10.672 26.083 0 319.853Dismissal Law Index 13079 0.070 0 0.082 0 0.167

United Kingdom

Obns. Mean Median Std. Devn. Min. Max.Number of patents 9849 7.503 4 12.172 1 286Number of citations 9849 40.881 18 68.422 0 1145Standard deviation of citations 7652 7.751 4.950 9.626 0 130.194Dismissal Law Index 9849 0.377 0.407 0.094 0.049 0.444

Germany

Obns. Mean Median Std. Devn. Min. Max.Number of patents 11402 17.221 9 23.286 1 349Number of citations 11402 76.343 35 113.174 0 1313Standard deviation of citations 10004 9.675 5.859 12.606 0 174.062Dismissal Law Index 11402 0.431 0.425 0.018 0.407 0.488

France

Obns. Mean Median Std. Devn. Min. Max.Number of patents 9625 7.299 4 11.055 1 250Number of citations 9625 33.398 15 51.078 0 747Standard deviation of citations 7361 7.285 4.295 9.977 0 163.613Dismissal Law Index 9625 0.700 0.746 0.148 0.281 0.782

Panel B: U.S. WARN Sample

Obns. Mean Median Std. Devn. Min. Max.Number of patents 13099 17.117 2 61.041 1 1622Number of citations 13099 173.697 24 706.825 0 21164Number of employees 11704 16147 2552 45635 0 876800Market-to-Book 10389 2.097 1.337 4.270 0.304 260.180Size 11655 5.565 5.763 2.583 -6.215 11.933

26

Table II: Cross-country tests: Country- and industry-level fixed effects panel regressions.

The OLS regressions in Panel A of the table below implement the following model:yct = βc + βt + β1 ∗DismissalLawsc,t−l + εctwhere yct is the natural logarithm of a measure of innovation from country c in year t. βc and βt denote country and applicationyear fixed effects, respectively. DismissalLawsc,t−l denotes the lth lag of the dismissal law index for country c (from Deakinet al., 2007). The sample in Panel A spans the years 1971–2002. Heteroscedasticity-consistent standard errors are given inparentheses.The OLS regressions in Panel B implement the following model:yict = tβj←i + tβc + βi + βc + βt + β1 ∗DismissalLawsc,t−l + βXict + εictwhere yict is the natural logarithm of a measure of innovation for the USPTO patent class i from country c applied for in yeart. tβj←i denotes a time trend for the industry (patent category) j to which patent class i belongs; tβc denotes a time trend forcountry c. βi, βc, βt denote patent class, country and application year fixed effects. Xict denotes the set of control variables.The sample in Panel B is for 1978–2002. Robust standard errors (clustered by country-year) are given in parentheses.***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively.

Panel A

(1) (2) (3) (4) (5) (6)Dependent Variable is Number of Number of Std. Dev. of Number of Number of Std. Dev. ofNat. Logarithm of Patents Citations Citations Patents Citations Citations

Dismissal Law Index(t-1) 0.349** 0.430** 0.567**(0.168) (0.182) (0.225)

Dismissal Law Index(t-2) 0.335** 0.314* 0.393*(0.159) (0.185) (0.225)

Country and Application Year FE X X X X X XObservations 128 128 128 124 124 124Adjusted R-squared 0.992 0.993 0.971 0.992 0.993 0.971

Panel B

(1) (2) (3) (4) (5) (6)Dependent Variable is Number of Number of Std. Dev. of Number of Number of Std. Dev. ofNat. Logarithm of Patents Citations Citations Patents Citations Citations

Dismissal Law Index(t-1) 1.844*** 4.628*** 1.056**(0.344) (1.258) (0.416)

Dismissal Law Index(t-2) 1.546*** 3.449*** 0.786**(0.420) (1.305) (0.386)

Other Control Variables employed: Creditor Rights Index, log of per capita GDP,log(Imports), log(Exports), Ratio of Value Added

Patent Class, Country, X X X X X XApplication Year FEPatent Category and X X X X X XCountry-Specific TrendsObservations 23,385 23,385 20,194 23,385 23,385 20,194Adjusted R-squared 0.836 0.825 0.679 0.836 0.825 0.679

27

Table III:Cross-country evidence: Tests accounting for potential endogeneity of dismissal laws.

The OLS regressions in Panel A implement the following model: yct = βc + βt + β1 ∗DismissalLawsc,t−l + βXct + εctwhere yct is the natural logarithm of a measure of innovation from country c in year t. βc and βt denote country and applicationyear fixed effects. DismissalLawsc,t−l denotes the lth lag of the dismissal law index for country c. Xct denotes the controlvariables. Heteroscedasticity-consistent standard errors are given in parentheses below the coefficients.The OLS regressions in Panel B implement the following model:yict = tβj←i + tβc + βi + βc + βt + β1 ∗DismissalLawsc,t−l + βXict + εictwhere yict is the natural logarithm of a measure of innovation for the USPTO patent class i from country c in year t. tβj←i

denotes a time trend for the industry (patent category) j to which patent class i belongs; tβc denotes a time trend for countryc. βi, βc, βt denote patent class, country and application year fixed effects. Xict denotes the control variables. Robust standarderrors (clustered by country-year) are given in parentheses. The sample in both panels covers the years 1978–2002. ***, **,and * denote significance at the 1%, 5%, and 10% levels, respectively.

Panel A

(1) (2) (3) (4) (5) (6)Dependent Variable is Number of Number of Std. Dev. of Number of Number of Std. Dev. ofNat. Logarithm of Patents Citations Citations Patents Citations Citations

Dismissal Law Index(t-1) 1.696*** 1.693*** 1.563**(0.258) (0.524) (0.607)

Dismissal Law Index(t-2) 1.739*** 1.572*** 1.208*(0.251) (0.540) (0.647)

Government 0.005 0.008 0.034 0.008 0.011 0.036(0.009) (0.025) (0.027) (0.009) (0.025) (0.027)

Real GDP Growth rate (%) 0.006 0.022 0.027 0.006 0.021 0.027(0.008) (0.020) (0.020) (0.008) (0.020) (0.020)

Recession Dummy -0.043 0.022 -0.067 -0.024 0.040 -0.053(0.048) (0.122) (0.118) (0.046) (0.120) (0.118)

Other Control Variables employed: Creditor Rights Index, log of per capita GDPCountry and Application Year FE X X X X X XObservations 100 100 100 100 100 100Adjusted R-squared 0.996 0.994 0.976 0.996 0.994 0.975

Panel B

(1) (2) (3) (4) (5) (6)Dependent Variable is Number of Number of Std. Dev. of Number of Number of Std. Dev. ofNat. Logarithm of Patents Citations Citations Patents Citations Citations

Dismissal Law Index(t-1) 1.730*** 4.240*** 0.791**(0.313) (1.143) (0.335)

Dismissal Law Index(t-2) 1.341*** 2.756** 0.432(0.404) (1.138) (0.300)

Government 0.008 0.025 0.021*** 0.012** 0.038** 0.024***(0.005) (0.016) (0.006) (0.006) (0.017) (0.006)

Real GDP Growth rate (%) 0.001 0.004 -0.002 0.002 0.009 -0.001(0.004) (0.010) (0.005) (0.005) (0.011) (0.005)

Recession Dummy -0.043* -0.124* -0.066*** -0.023 -0.081 -0.058***(0.024) (0.063) (0.022) (0.026) (0.066) (0.021)

Other Control Variables employed: Creditor Rights Index, log of per capita GDP,log(Imports), log(Exports), Ratio of Value Added

Patent Class, Country, X X X X X XApplication Year FEPatent Category and X X X X X XCountry TrendsObservations 23,385 23,385 20,194 23,385 23,385 20,194Adjusted R-squared 0.836 0.826 0.680 0.836 0.825 0.679

28

Table IV: Cross-CountryEvidence: Triple-difference tests controlling for all omitted variables at the country level.

The OLS regressions in Columns 1–6 below implement the following model:yict = tβj←i + βc,t + βi + β1 ·DismissalLawsc,t−1 · InnovationIntensityi,US + βXict + εictwhere yict is the natural logarithm of a measure of innovation for the USPTO patent class i from country c applied for in yeart. tβj←i denotes a time trend for the industry (patent category) j to which patent class i belongs; βc,t denotes the country-year fixed effects. βi denotes patent class fixed effects. DismissalLawsc,t−1 denotes the index of laws governing dismissal incountry c in year (t− 1). InnovationIntensityi,US is the number of R&D scientists and engineers per 1,000 employees in U.S.manufacturing companies for patent class i, averaged over the years 1983–1993. Xict denotes the set of the following controlvariables: log(Imports), log(Exports), Ratio of Value Added. The sample period in Columns 1–6 spans 1971–2002. Robuststandard errors (clustered by country-year) are given in parentheses.Tests in Columns 7 & 8 estimate the following regression model:yict, firms − yict, individuals = tβj←i + tβc + βi + βc + βt + β1 ∗DismissalLawsc,t−1 + βXict + εictIn addition to the control variables from Columns 1–6, we also use the Creditor Rights Index and the log of per capita GDP inColumns 7 & 8. The sample period is 1978–2002. Robust standard errors (clustered by country-year) are given in parentheses.***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively.

(1) (2) (3) (4) (5) (6) (7) (8)excluding U.S.

Dependent Variable is Nb. of Nb. of SD. of Nb. of Nb. of SD. of Innovation by FirmsNat. Logarithm of Patents Citations Citations Patents Citations Citations - Innovation by Individuals

Number of Number ofPatents Citations

Dismissal Law Index(t-1) 0.007*** 0.012*** 0.004*** 0.011*** 0.017*** 0.008**** InnovationIntensity (0.001) (0.002) (0.001) (0.002) (0.004) (0.003)

Dismissal Law Index(t-1) 0.674*** 2.404***(0.206) (0.756)

Control Variables X X X X X X X XCountry x Year FE X X X X X XCountry and Year FE X XCountry–Specific Trends X XPatent Category Trends X X X X X X X XPatent Class FE X X X X X X X XObservations 16,033 16,033 13,732 10,465 10,465 8,341 23,385 23,385Adjusted R-squared 0.830 0.826 0.652 0.674 0.681 0.528 0.735 0.641

29

Table V: Cross-country evidence: Dynamic effects.

The tests in the table below report the results examining the impact of the dismissal law change in the U.S. in 1989; the “controlgroup” is Germany, which did not experience such a law change in the sample period (from 1970-1995). The tests examine thepossibility of reverse causality by following Bertrand and Mullainathan (2003) in decomposing the federal change in dismissallaws in the U.S. into three separate time periods: Dismissal Law Change (-2,0) is a dummy that takes the value of one forthe years 1987-1989 for the U.S., zero otherwise; Dismissal Law Change (1,2) is a dummy that takes the value of one for theyears 1990-1991 for the U.S., zero otherwise; finally, Dismissal Law Change (≥3) is a dummy that takes the value of one forthe years 1992 and thereafter for the U.S., zero otherwise.The labor law index data is from Deakin et al. (2007). The sample period spans 1970–1995. Robust standard errors (clusteredby country-year) are given in parentheses. ***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively.

(1) (2) (3)Dependent Variable is Number of Number of Std. Dev. ofNat. Logarithm of Patents Citations Citations

Dismissal Law Change (-2,0) -0.138*** -0.039 0.033***(0.035) (0.034) (0.009)

Dismissal Law Change (1,2) 0.066* 0.214*** 0.187***(0.035) (0.037) (0.028)

Dismissal Law Change (≥3) 0.163*** 0.303*** 0.204***(0.035) (0.032) (0.019)

Patent Class, Country, X X XApplication Year FEObservations 19,621 19,621 18,281Adjusted R-squared 0.841 0.818 0.608

Table VI: Effect of dismissal laws vis-a-vis other dimensions of labor laws.

The OLS regressions below implement the following model:yict = tβj←i + tβc + βi + βc + βt + β1 ∗ lAc,t−1 + β2 ∗ lBc,t−1 + β3 ∗ lCc,t−1 + β4 ∗ lDc,t−1 + β5 ∗ lEc,t−1 + βXict + εictwhere yict is the natural logarithm of a measure of innovation for the USPTO patent class i from country c applied for inyear t. tβj←i denotes a time trend for the industry (patent category) j to which patent class i belongs; tβc denotes a timetrend for country c. βi, βc, βt denote patent class, country and application year fixed effects. β1 - β5 measure the impacton measures of innovation of the respective labor law for the five components of the Deakin et al. (2007) labor law index:Alternative employment contracts (lAc,t−1), Regulation of working time (lBc,t−1), Regulation of Dismissal / Dismissal LawIndex (lCc,t−1), Employee representation (lDc,t−1), and Industrial action (lEc,t−1). Xict denotes the set of control variables.The sample period is 1978–2002. Robust standard errors (clustered by country-year) are given in parentheses. ***, **, and *denote significance at the 1%, 5%, and 10% levels, respectively.

(1) (2) (3)Dependent Variable is Number of Number of Std. Dev. ofNat. Logarithm of Patents Citations Citations

Dismissal Law Index(t-1) 1.911*** 4.513*** 1.038***(0.346) (1.107) (0.383)

Regulation Of Working Time(t-1) -0.364 1.795** 0.899***(0.270) (0.761) (0.304)

Alternative Employment Contracts(t-1) -0.288** -0.227 -0.062(0.118) (0.285) (0.149)

Employee Representation(t-1) 0.441 -1.978** -0.841**(0.312) (0.863) (0.359)

Industrial Action(t-1) 0.321 1.197 -0.003(0.431) (1.408) (0.434)

Other Control Variables employed: Creditor Rights Index, log of per capita GDP,log(Imports), log(Exports), Ratio of Value Added

Patent Class, Country, X X XApplication Year FEPatent Category and X X XCountry TrendsObservations 23,385 23,385 20,194Adjusted R-squared 0.836 0.826 0.679

30

Tab

leV

II:

Wit

hin

-cou

ntr

yevid

ence:

Diff

ere

nce-i

n-d

iffere

nce

test

s.

The

regre

ssio

ns

bel

owim

ple

men

tth

efo

llow

ing

model

:yit

=βi

+βt

+β

1∗

(Over

100) i

,1987∗

(After1

988) t

+βX

it+ε i

t

βi

andβt

are

firm

and

yea

rfixed

effec

ts,

resp

ecti

vel

y.(Over

100) i

,1987

isa

dum

my

vari

able

takin

gth

eva

lue

of

one

inea

chyea

rif

afirm

has≥

100

emplo

yee

sin

1987,

and

zero

oth

erw

ise;

as,

for

agiv

enfirm

,th

isva

riable

does

not

vary

over

tim

e,it

seff

ect

issu

bsu

med

inth

efirm

dum

mie

s.(After1

988) t

isa

dum

my

takin

gth

eva

lue

of

one

aft

erth

epass

age

of

the

WA

RN

Act

(i.e

.th

eyea

rs1989-1

994);

this

coeffi

cien

tis

subsu

med

by

the

yea

rdum

mie

s.In

Colu

mns

1–4,

the

dep

enden

tva

riable

sare

the

natu

ral

logari

thm

of

pate

nts

,as

wel

las

the

natu

ral

logari

thm

of

cita

tions.

InC

olu

mns

5–8,

the

pate

nts

and

cita

tions

are

scale

dby

the

num

ber

of

emplo

yee

sb

efore

takin

gth

elo

g.

Pate

nt

data

isfr

om

the

NB

ER

Pate

nts

File

(Hall,

Jaff

eand

Tra

jten

ber

g,

2001).

Fir

m-l

evel

data

isfr

om

Com

pust

at.

The

sam

ple

per

iod

is1983–1994.

Robust

standard

erro

rs(c

lust

ered

at

the

firm

level

)are

giv

enin

pare

nth

eses

.***,

**,

and

*den

ote

signifi

cance

at

the

1%

,5%

,and

10%

level

s,re

spec

tivel

y.

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Dep

end

ent

vari

ab

leis

lnof

Pate

nts

Cit

ati

on

sP

ate

nts

Cit

ati

on

sP

ate

nts

/C

itati

on

s/

Pate

nts

/C

itati

on

s/

Em

plo

yee

sE

mp

loyee

sE

mp

loyee

sE

mp

loyee

s

(Over

100) i,1

987∗

(After1988) t

0.1

69***

0.4

47***

0.2

16***

0.4

77***

0.6

80***

0.9

78***

0.3

23***

0.5

80***

(0.0

47)

(0.0

81)

(0.0

52)

(0.1

03)

(0.1

34)

(0.1

34)

(0.0

92)

(0.1

14)

Siz

e0.1

79***

0.1

41***

-0.3

94***

-0.4

37***

(0.0

25)

(0.0

38)

(0.0

27)

(0.0

38)

Mark

et-t

o-B

ook

-0.0

02

-0.0

01

-0.0

05

-0.0

03

(0.0

04)

(0.0

07)

(0.0

06)

(0.0

08)

Fir

man

dA

pp

lica

tion

Yea

rF

EX

XX

XX

XX

XO

bse

rvati

on

s13,0

67

13,0

67

10,2

97

10,2

97

11,6

93

11,6

93

10,2

29

10,2

29

Ad

just

edR

-squ

are

d0.8

51

0.7

36

0.8

68

0.7

54

0.9

04

0.8

24

0.9

24

0.8

29

31

Table VIII: Within-country evidence: Regression discontinuity tests.

The regressions below implement the following model:yit = βi + βt + β1 ∗ (Over100)i,1987 ∗ (After1988)t + βXit + εitβi and βt are firm and year fixed effects, respectively. Across all three panels, the dependent variables are (the log of) patentsand citations, as well as (the log of) patents and citations scaled by the number of employees. In addition, we also employas dependent variable Ind(Empi,t − Empi,t−1 < 0), a binary variable taking on a value of one in case of a net employmentreduction in firm i from year t− 1 to year t. (Over100)i,1987 is a dummy variable taking the value of one in each year if a givenfirm has ≥ 100 employees in 1987, and zero otherwise; as, for a given firm, this variable does not vary over time, its effect issubsumed in the firm dummies. (After1988)t is a dummy taking the value of one after the passage of the WARN Act (i.e. theyears 1989-1994); this coefficient is subsumed by the year dummies. In Panel A, the sample is restricted to firms whose 1987employment is just below or just above the relevant WARN cutoff, i.e. firms with employment between 90 and 110 employees.In Panel B, the sample is restricted to firms whose 1987 employment is between 40 and 60 employees. Finally, in Panel C,only firms whose employment in the year 1987 is between 140 and 160 are included in the sample.The sample period is 1983–1994. Robust standard errors (clustered at the firm level) are given in parentheses. ***, **, and *denote significance at the 1%, 5%, and 10% levels, respectively.

Panel A: 90 ≤ Employmenti,1987 < 110

(1) (2) (3) (4) (5)Dependent Variable is Ind(Empi,t − Empi,t−1 < 0) Ln(Patents) Ln(Citations) Ln(Patents / Ln(Citations /

Employees) Employees)

(Over100)i,1987 ∗ (After1988)t -0.293** 0.370** 0.606*** 0.599 0.724*

(0.114) (0.146) (0.217) (0.444) (0.432)(Over100)i,1987 0.283***

(0.083)Firm FE X X X XApplication Year FE X X X X XNumber of Firms 117 239 239 219 219Observations 435 916 916 765 765Adjusted R-squared 0.027 0.828 0.693 0.649 0.681

Panel B: 40 ≤ Employmenti,1987 < 60

(1) (2) (3) (4) (5)Dependent Variable is Ind(Empi,t − Empi,t−1 < 0) Ln(Patents) Ln(Citations) Ln(Patents / Ln(Citations /

Employees) Employees)

(Over50)i,1987 ∗ (After1988)t -0.076 0.176 0.633 -0.468* 0.191

(0.209) (0.152) (0.598) (0.254) (0.476)(Over50)i,1987 -0.030

(0.147)Firm FE X X X XApplication Year FE X X X X XNumber of Firms 39 77 77 70 70Observations 115 259 259 230 230Adjusted R-squared 0.039 0.188 0.348 0.633 0.492

Panel C: 140 ≤ Employmenti,1987 < 160

(1) (2) (3) (4) (5)Dependent Variable is Ind(Empi,t − Empi,t−1 < 0) Ln(Patents) Ln(Citations) Ln(Patents / Ln(Citations /

Employees) Employees)

(Over150)i,1987 ∗ (After1988)t 0.104 -0.115 0.201 -0.164 0.033

(0.159) (0.357) (0.451) (0.592) (0.717)(Over150)i,1987 -0.124

(0.181)Firm FE X X X XApplication Year FE X X X X XNumber of Firms 21 40 40 37 37Observations 88 170 170 155 155Adjusted R-squared 0.073 0.490 0.633 0.447 0.516

32