Exercise 2

Christoph Huber

1
Shareholder Litigation and Corporate Innovation
• Chen Lin, Sibo Liu, Gustavo Manso (2021) Shareholder Litigation and Corporate Innovation.
Management Science 67(6):3346-3367. https://doi.org/10.1287/mnsc.2020.3626

Abstract
We investigate whether and to what extent shareholder litigation shapes corporate innovation by examining the staggered adoption of
universal demand laws in 23 states from 1989 to 2005. These laws impose obstacles against shareholders filing derivative lawsuits,
thereby significantly reducing managers’ litigation risk. Using a difference-in-differences design and a matched sample, we find that,
following the passage of the laws, firms invested more in research and development, produced more patents in new technological
classes and more patents based on new knowledge, generated more patents with significant impacts, and achieved higher patent
value. Our findings suggest that the external pressure imposed by shareholder litigation discourages managers from engaging in
explorative innovation activities.

2
Shareholder Litigation and Corporate Innovation
• Chen Lin, Sibo Liu, Gustavo Manso (2021) Shareholder Litigation and Corporate Innovation.
Management Science 67(6):3346-3367. https://doi.org/10.1287/mnsc.2020.3626
• Goal of this exercise:
▪ reproduce (and thereby understand) the original methods / results

3
Shareholder Litigation and Corporate Innovation
What is the causal relationship of interest?
Two competing hypotheses:

1. disciplining hypothesis
• the threat of shareholder litigation “disciplines” a manager’s behavior and stimulates
corporate innovation
• shareholder litigation → higher corporate innovation

2. pressure hypothesis
• shareholder litigation stifles corporate innovation
• shareholder litigation → lower corporate innovation

4
Shareholder Litigation and Corporate Innovation
What is the paper’s identification strategy?
→ Difference-in-differences design

• p. 3346: “Our empirical investigation relies on the staggered adoption of the universal
demand (UD) laws that generate inter-temporal reductions in litigation risk among firms.
Between 1989 and 2005, 23 states passed the laws that raise the barriers for shareholders to
file derivative lawsuits against a firm’s management.”

• “… Although the UD laws may not be randomly assigned, the staggered adoption process
across states enables us to apply a difference-in-differences approach to understand the
relationship between shareholder litigation and corporate innovation.”

5
Data
• Treated firms: “The treated firms are those incorporated in states that adopted the laws
between 1976 and 2006.” (p. 3347)
• Non-treated firms (control group): a group of control firms that are headquartered in the same
state as the corresponding treated firms but are incorporated in states without the laws

→ compare the firms within the same headquarter state but subject to different legal
institutions regarding shareholder litigation mandated by laws at the incorporation state level
→ matched sample containing 19,096 firm-year observations of U.S. public firms

• “We use a propensity score matching approach to construct a sample of treated firms and
control firms that shared similar traits before the laws were implemented.”

6
Propensity Score Matching (PSM) in DiD
• Goal: Improve comparability by selecting a better control group.
▪ In DiD, we compare treated and control groups over time. If these groups differ
systematically at the start, the comparison might be biased. PSM helps create a more
comparable control group.
• Propensity Score: Probability of treatment based on observed characteristics (e.g., firm size,
performance); estimated using logistic regression
• Matching: Pair treated firms with non-treated firms that have similar propensity scores.
• Why Combine with DiD? PSM balances observed traits; DiD removes time-invariant biases.

• Limitations: PSM only balances observed characteristics!

7
Descriptive Statistics
Step 1: Load the data set

• In R, the package haven with the function

Step 2: Get a first impression of the data

• Descriptive statistics
• Potentially plot the data in some reasonable
way
▪ e.g., timing of UD Law adoption

8
Descriptive Statistics
Step 1: Load the data set
1 library(haven)
2 data <- read_stata("Exercise_2_materials/Lin_et_al_2020_data.dta")

Descriptives:
1 library(psych)
2 describe(data[, c("law", "xrdf_atw", "ln_patent_count")])[, c("n", "mean", "sd")]
n mean sd
law 19096 0.21 0.41
xrdf_atw 19096 0.03 0.09
ln_patent_count 19096 0.43 1.00
1 library(tidyverse)
2 data %>%
3 summarise(
4 n = n(),
5 law_mean = mean(law),
6 sd_mean = sd(law)
7 )
# A tibble: 1 × 3
n law_mean sd_mean
<int> <dbl> <dbl>
1 19096 0.211 0.408

9
Difference-in-differences specification
p. 3351: Our difference-in-differences specification is as follows:
Innovationi,t = α + βLawi,t + γXi,t + θi + δi,t + εi,t

where:
Innovationi,t … innovation measure gauged by several proxies
Lawi,t … equals one if the incorporation state of the firm has a UD law
β … main coefficient of interest for identifying the effect of the laws
θi … firm fixed effects that capture the firm-level time-invariant effects
δi,t … headquarter state by year fixed effects that pick up regional trends
Xi,t … lagged firm-level at- tributes, including firm size, leverage, book-to-market ratio, firm
age, and profitability
10
Difference-in-differences specification
p. 3351: Our difference-in-differences specification is as follows:
Innovationi,t = α + βLawi,t + γXi,t + θi + δi,t + εi,t

Note the similarity with the two-way fixed effects specification from the lecture:
Investmenti,t = γi + λt + δT axReformi,t + εi,t

• two fixed-effects terms and one term for the treatment on the firm-year level
(subscripts i and t)

11
Let’s implement it in R
1 libray(fixest)
2
3 model <- feols(... ~ ... | ... + ...,
4 cluster = ~...,
5 data = data)

12
Table 5 - Regression Code
1 library(fixest)
2
3 model1 <- feols(xrdf_atw ~ law | gvkey_num + state_year,
4 cluster = ~incorp_hist_id,
5 data = data)

1 model2 <- feols(xrdf_atw ~ law + l1_ln_atw + l1_levw + l1_mbw + l1_ln_age +

13
Table 5 - Output
1 library(modelsummary)
2 msummary(list(model1, model2),
3 stars = c("*" = 0.10, "**" = 0.05, "***" = 0.01),
4 coef_omit = "l1",
5 gof_omit = "AIC|BIC|RMSE|Within")

(1) (2)
law 0.003** 0.002**
(0.001) (0.001)
Num.Obs. 19096 19096
R2 0.678 0.701
R2 Adj. 0.641 0.667
Std.Errors by: incorp_hist_id by: incorp_hist_id
FE: gvkey_num X X
FE: state_year X X
* p < 0.1, ** p < 0.05, *** p < 0.01

14
Dynamic vs. Static Effects
• Static effect: So far, in the lecture on DiD, we have considered the average treatment effect at
a particular post-treatment period (e.g., the year a�er treatment for all treated units).
→ here, you might have noticed that this is different: we have several periods before and
several periods a�er the treatment comes into effect

• Dynamic effect: You look at the treatment effect at multiple time points relative to the
treatment event.
▪ This o�en involves an visualization showing the treatment effect at various times relative
to treatment (e.g., years before and a�er the treatment event).

15
Dynamic vs. Static Effects - Formulas
• Static effect:
▪ yit = αi + λt + β ⋅ Di,t + γ ⋅ Xi , t + ϵi,t
• Dynamic effect:
▪ yi,t = αi + λt + ∑k δk ⋅ 1(t−_treati =k) + γ ⋅ Xi,t + ϵi,t
▪ can be used for identifying pre-trends before the treatment
◦ see Table 9 in Lin et al. (2020):
3
◦ Innovationi,t = α + ∑k=−3 β k Law+ γXi,t + θi + δi,t + εi,t
▪ can be used for measuring the dynamics of the treatment effect

16
Dynamic Effect - R (Table 9, Column 1)
1 model <- feols(xrdf_atw ~ d_3 + d_2 + d_1 + d0 + d1 + d2 + d3x | gvkey_num + state_year,
2 cluster = ~incorp_hist_id,
3 data = data)
4 summary(model)
OLS estimation, Dep. Var.: xrdf_atw
Observations: 19,096
Fixed-effects: gvkey_num: 942, state_year: 1,009
Standard-errors: Clustered (incorp_hist_id)
Estimate Std. Error t value Pr(>|t|)
d_3 0.001330 0.000985 1.350130 0.185402
d_2 -0.002712 0.001946 -1.393960 0.171875
d_1 0.001640 0.002608 0.628809 0.533444
d0 0.002071 0.001979 1.046292 0.302401
d1 -0.000156 0.001979 -0.078725 0.937687
d2 0.004736 0.002471 1.916407 0.063279 .
d3x 0.003656 0.001786 2.046663 0.048046 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
RMSE: 0.049495 Adj. R2: 0.641203
Within R2: 3.294e-4

• none of the pre-treatment dummy variables is statistically significant → no pre-trends (good