corporate performance and cultureconference.iza.org/conference_files/data_2018/pasch_s26796.pdf ·...

53
1 Corporate Culture and Industry-Fit: A Text Mining Approach * Stefan Pasch March 27, 2018 Abstract Even though several studies point out the importance of choosing a corporate culture that fits a firm’s business environment, corporate cultures usually vary considerably among firms within the same industry. This paper tests whether it is actually beneficial for a firm to have a culture that differs from their industry standard. To do so, I web scrape over 550,000 employee reviews from Glassdoor.com, a career community website, and apply text analysis techniques to measure corporate culture based on these reviews. Firms that differ strongly from the average culture of their industry show worse firm performance, supporting the hypothesis that a culture should fit to its business environment. Moreover, I find that suboptimal culture choices can be partly explained by CEO characteristics, while regional culture only plays a minor role. This suggests that in addition to management practices (Bloom and Van Reenen 2007) choosing an appropriate culture can be an important channel through which managers can influence firm performance. Keywords: Corporate Culture, Firm Performance, Industry-Fit, Person-Fit, Managers, Text Mining, Big Data JFL Classification: L22, L25, M12 * I thank Iwan Barankay, Guido Friebel, Mitchell Hoffmann, Christopher Koch, Bentley MacLeod, Rajesh Ramachandran, Jean Tirole, and Elisa Wirsching as well as seminar participants in Frankfurt for helpful discussions and comments Goethe University Frankfurt. E-Mail: [email protected]

Upload: others

Post on 19-Apr-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

1

Corporate Culture and Industry-Fit:

A Text Mining Approach*

Stefan Pasch†

March 27, 2018

Abstract

Even though several studies point out the importance of choosing a corporate

culture that fits a firm’s business environment, corporate cultures usually vary

considerably among firms within the same industry. This paper tests whether it is

actually beneficial for a firm to have a culture that differs from their industry

standard. To do so, I web scrape over 550,000 employee reviews from

Glassdoor.com, a career community website, and apply text analysis techniques to

measure corporate culture based on these reviews. Firms that differ strongly from

the average culture of their industry show worse firm performance, supporting the

hypothesis that a culture should fit to its business environment. Moreover, I find

that suboptimal culture choices can be partly explained by CEO characteristics,

while regional culture only plays a minor role. This suggests that in addition to

management practices (Bloom and Van Reenen 2007) choosing an appropriate

culture can be an important channel through which managers can influence firm

performance.

Keywords: Corporate Culture, Firm Performance, Industry-Fit, Person-Fit, Managers,

Text Mining, Big Data

JFL Classification: L22, L25, M12

* I thank Iwan Barankay, Guido Friebel, Mitchell Hoffmann, Christopher Koch, Bentley MacLeod, Rajesh

Ramachandran, Jean Tirole, and Elisa Wirsching as well as seminar participants in Frankfurt for helpful discussions

and comments

† Goethe University Frankfurt. E-Mail: [email protected]

Page 2: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

1

1. Introduction

One established finding in the literature about corporate culture is that industry characteristics are

an important determinant for a firm's corporate culture choice and that cultures differ more

strongly between industries than within industries (Chatman and Jehn, 1994; Gordon, 1991).

However, it remains debatable whether a firm benefits from conforming to their industry standard

or gains more by having a corporate culture that differs from their competitors. On the one hand,

a firm that offers a distinct culture might benefit, because it facilitates the matching between

workers and firms (Kosfeld and Von Siemens, 2011). Chatman (1991) argues that workers self-

select into firms with similar values and found workers with a close fit to their firm's value to be

more satisfied and to intend a longer employment at their firm (person-fit hypothesis). On the

other hand, when an industry culture emerges as an optimizing reaction to the business

environment, differentiating corporate culture from the average industry culture might reflect a

suboptimal HR management and poor industry knowledge (industry-fit hypothesis).

Therefore, this paper analyzes the association between deviation from the average

industry culture and firm performance. To do so, I web scrape over 550,000 employee reviews

from 2008 to 2017 from Glassdoor.com, a career community website, which allows me to

construct a sufficiently large sample to analyze within industry variation. The study is focused on

the United States and therefore only reviews from the U.S. are included. As a measure of

corporate culture I rely on the Organizational Culture Profile (OCP) (O'Reilly, Chatman and

Caldwell, 1991), which groups various values and attributes related to corporate culture in seven

dimensions: Innovation, stability, respect for people, outcome orientation, attention to detail,

team orientation, and aggressiveness. For each dimension of the OCP I create a master text using

the WordNet library by finding synonyms and hyponyms. To construct measures of corporate

culture, I calculate the text similarities between the master texts and the employee reviews from

Glassdoor. These measures of culture are matched with data on firm characteristics and

performance from COMPUSTAT.

At first, I scrutinize the general association between culture and firm performance and

find a concave relationship with firm performance for 6 out of the 7 dimensions of culture. This

indicates that different to management practices (Bloom, Sadun, and van Reenen 2016) culture

Page 3: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

2

does not function as a technology, which affects performance linearly. The absence of a linear

effect hints that the association between culture and performance indeed also depends on

additional factors like the fit to the business environment. In fact, with regard to industry

dynamics I find that firms that differ from their industry's1 average corporate culture are

associated with significantly worse firm performance, which supports the industry-fit hypothesis.

The effect is also economically important with a divergence of one standard deviation to the

average culture being associated with over $250k or 11% lower net income. Therefore, this paper

adds to the discussion of how corporate culture can be linked with firm performance (e.g. Guiso,

Sapienza, and Zingales 2015, Sørensen 2002) by outlining the importance of fit to industry

characteristics.

Moreover, this paper contributes to the literature on organization capital and its external

and internal influences (Dessein and Prat, 2017). Bloom and van Reenen (2010) find that the

adoption of different management practices is highly correlated with firm performance. These

management practices include: lean manufacturing techniques, performance setting, target

tracking and talent acquisition. My results suggest that corporate culture might be a further

channel for how management practices affect firm performance. Accordingly, I find that CEO

tenure and whether a firm had a recent switch in CEO can partly explain why firms differ from

their average industry culture. This is in line with Bertrand and Schoar (2013), who identify

strong CEO fixed-effects and Berson, Oreg, and Dvir (2008) who establish a link between CEO

values and corporate culture. Contrariwise, I find that regional differences only play a minor role.

This paper also adds to a growing literature that scrutinizes advances in big data, online

markets, and automated text analysis approaches. Hansen and McMahon (2016: S114-S115)

“believe that the approach of using computational linguistics to create measures of

communication from large databases of text has broader applications […] and can help bringing

economics into the increasingly important world of Big Data”. They apply computational

linguistic tools on statements released by the FOMC to analyze its effect on market and real

economic variables. Other applications of text mining include political attention (Quinn et al.

1 I use the term industry loosely to refer to different business fields. I apply different classifications of the Global

Industry Classification Standard (GICS) to classify different business fields. Note that industry is also a classification

of the GICS. However, when I refer to industry as a GICS classification rather than a loose term for fields of

business, I mention this explicitly.

Page 4: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

3

2010), patents (Yoon and Park, 2014), and helpfulness of online reviews (Ghose and Panagiotis,

2011). Reviews from Glassdoor have been previously scrutinized to link employee satisfaction

with firm performance (Moniz, 2015), CEO personality (O’Reilly et al. 2014), and to link

corporate culture with shareholder governance (Popadak 2013). All studies point out that online

reviews contain valuable information about corporate culture. In addition, I am able to identify

industry specific cultures based on online employee reviews, similar to Chatman and Jehn (1994),

who identify differences between industries by surveying firm members.

The remainder of this paper is structured as follows. Section 2 provides a literature review

about corporate culture. Section 3 introduces the datasets and the measurement of corporate

culture. The empirical results are presented in Section 4. Section 5 analyzes the influence of

regional differences. Section 6 tests the robustness of the results. Section 7 concludes.

2. Literature Review

Although more than 4600 articles have been generated on the topic of corporate culture (Hartnell,

Ou, and Kinicki 2011) it remains a vague concept with varying definitions and interpretations.

For Schein (1992) culture represents the basic assumptions within an organizations that provide

employees guidelines on how to behave in situations without formal rules. However, according to

this view culture is mostly tacit, implicit and hard to measure. Crémer (1993) describes corporate

culture as an unspoken code of communication within an organization that enhances coordination

between members of an organization. According to Kreps (1990) corporate culture serves as a

reputation and coordination device to overcome multiple equilibria problems.

In this paper I rely on O'Reilly and Chatman’s (1996: 160) definition of corporate culture

as “a system of shared values (that define what is important) and norms that define appropriate

attitudes and behaviors for organizational members (how to feel and behave)”. The main

advantage of adopting the view of culture as a collection of values is that it allows to detect

industry-specific standards of culture, because values differ considerably between industries

(Chatman and Jehn 1994). According to Guiso, Sapienza, and Zingales (2015) there are two

further advantages in applying this definition of corporate culture: First, it fits to common models

of culture from neoclassical economic theory (Guiso, Sapienza, and Zingales 2008). Second, it

Page 5: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

4

allows measuring corporate culture and hence to link corporate culture with firm performance

empirically.

In fact, various studies could establish an association between corporate culture and firm

performance. Guiso, Sapienza, and Zingales (2015) find that firms with a culture that is perceived

as trustworthy and ethical show stronger performance. Moreover, firms with a strong culture tend

to outperform those firms with a weaker culture (Gordon and DiTomaso 1992; Burt et al. 1994),

where “strong” culture describes that there are certain values and norms that are “widely shared

and strongly held throughout the organization” (O'Reilly and Chatman 1996: 166). In line with

this, Sørensen (2002) finds that a strong culture leads to a more stable firm performance.

However, there are also findings suggesting that there is no “one-size fits all” culture and

that the business environment is highly important for the right culture choice. Sørensen (2002)

reports that the positive effect of corporate culture on performance stability decreases when the

firm operates in a volatile industry, because a strong culture might hinder the flexibility of a

company, which is important in a volatile environment. Similarly, Chatman et al. (2014) suggest

that a culture of adaptability is particularly important for high-technology firms. On the contrary,

healthcare organizations, for which failure avoidance is particularly important, should rely on a

culture that outlines safety (Nieva and Sorra 2003). The hypothesis that a culture should fit to the

corresponding business environment is also supported by Heskett and Kotter (1992), who analyze

207 firms from 22 different industries and only find a weak correlation between a culture strength

and firm performance on the long-run. For a subgroup of 22 companies they conduct more in-

depth investigation and find that the companies with an appropriate culture for their environment

outperformed their competitors.

Accordingly, several studies show that industry characteristics are an important

determinant for a firm’s corporate culture choice: Chatman and Jehn (1994) analyze cultures at

four different industries and find that organizational culture varies more across industries than

within. Similarly, Lee and Yu (2004) detect distinct values that characterize an industry. For

instance, insurance firms tend to be more task oriented, while hospitals are significantly more

team oriented. Nevertheless, the culture of firms usually also vary considerably within industries

(Chatman 1991; Chatman and Jehn 1994). For example, Sheridan (1992) finds that the corporate

cultures of six public accounting firms varied significantly among each other. The industry-fit

Page 6: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

5

hypothesis suggests that differences in culture between industries arise, because different industry

characteristics require different corporate cultures.

In contrast to the industry-fit hypothesis there is a branch of literature on the matching

between workers and firms, which suggests that there are advantages of offering a distinct

culture. Chatman (1991) argues that workers self-select into firms with similar values and finds

workers with a close fit to their firm’s values to be more satisfied and to intend a longer

employment at their firm. This result indicates that a firm might benefit from deviating from its

industry’s standard culture. Offering a culture that is different to competitors’ culture allows

workers to self-select into this firm based on their preferences. In fact, Rivera (2012) surveyed

hiring committees of service firms and found that concerns about cultural fit often outweighed

concerns about candidates’ skills. Related to the person-fit hypothesis, Kosfeld and Siemens

(2011) set up a model, where workers have heterogeneous social preferences for cooperation, and

show that there always exists a separating equilibrium in which workers self-select into firms

based on their preferences.

The industry-fit and the person-fit hypothesis give opposing suggestions about, whether a

deviation from the standard industry culture is beneficial or detrimental. To the best of my

knowledge, this question has so far not been analyzed empirically in a large-scale study that

includes multiple industries and multiple firms for each industry.

3. Data

In order to analyze the relationship between firms’ divergence from industry-specific culture and

their economic performance I match online employee reviews from a career community website

with performance data from COMPUSTAT. In the analysis I only consider firms that are

included in the COMPUSTAT database and I only analyze reviews written by employees that are

employed in the United States. Considering only the United States has two major advantages:

First, it allows a matching between the employee reviews with a consistent dataset for

performance measures and firm characteristics. Secondly, this study is focused on industry

variation in culture. As differences in industry culture might confound with national cultures,

Page 7: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

6

including different countries might bias the results. Naturally, also regional cultures differ within

the U.S. In Section 5 I scrutinize the effect of regional culture in more detail.

The COMPUSTAT database contains various performance measures as EBIT, EBITDA,

paid dividends, and the net income. In my analysis I rely on EBITDA (Earnings before interest,

taxes, depreciation and amortization) and the net income as performance measures. I scrutinize

these particular performance measures as the net income shows the overall profitability of a firm,

while the EBITDA abstracts from potential biases in firm performances associated to taxation

and accounting strategies. Additionally, I collect data on the number of employees, and staff

expenses. Moreover, I utilize the sector-, industry-, and sub-industry classifications from

COMPUSTAT that are based on the Global Industry Classification Standard (GICS) to define

different fields of business.

My measures for corporate culture are generated by employee reviews from

Glassdoor.com, an employee review site. Prior studies on corporate culture usually conduct

surveys to measure culture. Relying on online reviews instead of surveys has two main

advantages: First, it allows gathering a large sample of firms and makes it possible to analyze

between and within industry variation of culture. Second, it includes a more diverse set of

respondents. Surveys usually include executives and high educated participants; e.g. business

school alumni (Chatman et al. 2014), top managers (Lee and Yu 2004) or top executives, who

measure the culture of competitors (Kotter and Heskett 1992, Sørensen 2002). However, top

executives might describe their or their competitor’s culture in the way it is advertised and not

how it is actually perceived by employees. Yet, advertised values might not reflect the actual

culture. For instance, Guiso, Sapienza, and Zingales (2015) find that advertised values are not

very important compared to perceived values.

On Glassdoor employees and former employees can anonymously review their employer

by giving them an overall rating (“star rating”), but also by rating them in the following

dimensions: Work/Life balance, culture and values, career opportunities, compensation and

benefits, and the performance of the senior management. These star ratings range from 1 to 5,

where a rating of 5 corresponds to the highest satisfaction level. In addition to these ratings,

employees can describe in a free text what they do and do not like about working at their

employer, and they can give an advice to the senior management about what to improve in the

company. Figure AF1 in the Appendix shows two examples of Glassdoor reviews. Moreover,

Page 8: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

7

employees can self-report their salary and job candidates can review their experiences during

their job interview. I particularly rely on data from Glassdoor.com, though there are numerous

similar career community websites, for several reasons:

First, the website is the largest of its kind and therefore benefits from a very diverse

audience (Popadak 2013; Moniz 2015). To assess the characteristics of users, I retrieve web

traffic statistics from Quantcast.com similar to Moniz (2015). Quantcast is specialized to measure

the audience of a website by tracking cookies and clickstream of users. This allows Quantcast to

estimate the demographical composition of websites’ users, while anonymity of users is retained.

According to Quantcast, Glassdoor has 37.6 million unique users. 87% of all web traffic is

generated from the United States, which supports the approach of this paper to exclusively

analyze the U.S. market. Table 1 reports descriptive statistics on the profile of Glassdoor visitors.

As it can be seen, users are fairly distributed across various categories as gender, income,

education, and ethnicity. Quantcast also provides an index that shows, how these demographics

correspond to the general internet population, where an index of 100 represents an exact

representation and an index below 100 an under- and above 100 an overrepresentation. Naturally,

users below 18 and above 65 years are underrepresented, as they are predominantly not part of

the workforce. Furthermore, high- educated users are overrepresented, which suggests that they

are making more use of modern job searching tools. According to Quantcast, Caucasians are

under-, while ethnic minorities,

especially Asians, are

overrepresented. The reason for this

might be that ethnic minorities need

to engage more strongly in job

searching, because they are

discriminated in the hiring process.

For instance, Betrand and

Mullainathan (2004) find that

resumes with White-sounding names

are 50 percent more likely to receive

callbacks for interviews than resumes Figure 1: Distribution of Overall Ratings

Page 9: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

8

with African-American sounding names.

TABLE 1 ABOUT HERE

Another important reason for relying exclusively on Glassdoor data is its data integrity

that tries to prevent fake reviews (Moniz 2015). If users or companies suspect fake reviews they

can report those posts and editors from Glassdoor will check their validity. Indeed, roughly 15%

of reviews are rejected or deleted because they violate the guidelines (Moniz 2015). To present

further evidence on the validity of the reviews I conduct the following robustness check: As you

would expect that fake reviews either give overwhelmingly positive or negative ratings, I check

how many of the reviewers gave their employer either the highest or the lowest rating in all of the

available categories in the star rating. I find that of the 574,256 reviews I gathered; only 3.8%

gave the highest rating in all categories and 2.3% the lowest rating in all categories. Moreover,

most raters give an overall rating of 3 or 4 with an average rating of 3.32. Figure 1 shows the

distribution of the overall ratings, which reflect general satisfaction with employers. This

provides evidence that raters in fact give sophisticated reviews and do not simply use Glassdoor

as a tool to show frustration and anger about their employer.

A third advantage for using Glassdoor data is the rich content of the reviews. On the one

hand, the free texts from the reviews allow extracting how corporate culture is perceived: Which

values are outlined by many employees, and how are they described? On the other hand, the star

ratings on various dimensions allow to measure the general satisfaction of employees and in turn

to disentangle the design of corporate culture from the overall feeling of comfort. Furthermore,

reviews contain exact dates and location, which allows me to construct a panel dataset as well as

to analyze regional effects.

As this study is focused on the U.S., I extract only those reviews that have been generated

by employees, who work or used to work, in a location in the U. S. Interns and part-time workers

are excluded. Finally, I only web scrape the reviews from those firms, which are in the

COMPUSTAT database and have overall at least 100 reviews2 from workers employed in the

U.S. This leaves me with a dataset of 772 firms and 4930 firm-year observations. As Glassdoor

was founded in 2007, I analyze the time span 2008-2017.

2 Popadak (2013) only includes observations with at least 100 reviews. Therefore, only firms that have overall at

least 100 reviews are included. Section 6 tests several cutoff levels for firm-year observations.

Page 10: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

9

Yet, some of these firm-year observations are generated by only very few reviews,

especially for the early years of the sample, as Glassdoor became more popular over the years.

This could potentially cause a bias, if observations that are generated by a small set of reviews,

tend to show more extreme culture scores. Either any reviewer barely writes about a certain

culture dimension or the culture score for this dimension is very large due to one review with a

strong emphasize on a particular dimension of culture. Finding a cutoff level, which determines

the minimum amount of reviews that are required to include a firm-year observation in the study,

involves a tradeoff: One the one hand, culture scores that are generated by only few reviews

might not reflect the actual corporate culture. On the other hand, a low cutoff level includes more

observations (also those generated by few reviews), which is of particular importance for this

study. To determine an industry standard of culture and to analyze within industry variation a

sufficient sample of firms for each industry is required. Taking this tradeoff into account, I first

drop those observations that are generated by less than 30 reviews. In Chapter 6 I show that my

results do not rely on this cutoff point. Moreover, my main results also hold when imputing

values for insufficient observations, as omitting these observations could lead to a sample

selection bias.

To construct measures of corporate culture based on employee reviews I follow the text

similarity approach of Popadak (2013). First, I construct master texts for corporate culture based

on the organizational culture profile (OCP) (O'Reilly, Chatman, and Caldwell 1991). To form the

OCP, O'Reilly, Chatman, and Caldwell (1991) consider several values and attributes related to

corporate culture and analyze how they are interrelated. By applying principal component

analysis they form seven dimensions of corporate culture: Innovation, stability, respect for

people, outcome orientation, attention to detail, team orientation, and aggressiveness. AT1 lists

the seven dimensions of the OCP and shows the associated attributes. For each of these seven

dimensions I create a master text by finding synonyms and hyponyms3 of the related values and

attributes using the WordNet library. As it is common for computational linguistic techniques,

stop words (for example, “a”, “and”, “the”) are removed and the remaining words are stemmed,

3 A Hyponym is “a word of more specific meaning than a general or superordinate term applicable to it. For

example, spoon is a hyponym of cutlery.“ (Oxford Dictionary Online 2018)

Page 11: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

10

which reduces words to their linguistic word stem (“manager” and “management” become

“manag”). After this cleaning process, I calculate the text similarity between the employee

reviews and these master texts. The text similarity measure compares how often words from the

master text occur in the employee reviews relative to the overall length of the reviews. The

measures for corporate culture in the empirical part are generated by the average text similarity

for a firm in each year for each dimension of culture from the OCP. For better interpretation these

measures are transferred into Z-scores, ranging from -3 to 8.47. An obvious caveat of this text

similarity approach is that it does not take into account whether words are used in positive or

negative contexts. In Chapter 6 I address this concern and show that the main results are not

sensitive to negations.

Alternatively, when applying computational linguistic techniques, economists frequently

rely on topic models as the Latent Dirichlet Allocation (LDA) (Blei, Ng, and Jordan 2003). The

LDA is a flexible algorithm, which automatically finds relevant topics in unstructured data due to

repeated co-occurrences of certain words. For instance, Hansen and McMahon (2016) apply LDA

on information released by the Federal Open Market Committee (FOMC) to detect relevant

topics discussed by the FOMC. However, I do not rely on LDA in my main analysis for a

particular reason: Applying LDA is suitable for detecting certain word groups and topics of a

particular corpus of texts. Yet, employee reviews also contain information and topics that are not

mainly about corporate culture such as: compensation and benefits, description of the firm and its

industry, position of the employee within the firm. Figure AF2 gives some examples of word

clouds that were generated by applying an LDA, with 15 topics and 30 words each, on the

employee reviews from Glassdoor. As it can be seen, most word groups are dominated by very

common word stems related to employment such as “manag”, “work” or “company”, which

makes it difficult to interpret these topics as certain styles or characteristics of corporate culture.

A potential solution would be to remove these common words from the corpus. Yet, there is no

clear decision rule for which words to be removed. In fact, eliminating words from the corpus

until the word groups from LDA become informative about corporate culture appears artificial

and would most likely bias the results. On the other hand, the dictionary method allows extracting

Page 12: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

11

only those words that are related to corporate culture and also to form and analyze different styles

of culture.4

Moreover, the topic analysis shows that word clouds also cover words that reflect the

general satisfaction of employees (“great”, “good”), compensation and benefits (“pay”,

“benefits”, “contract”), and work life balance (“time”, “life”, “hour”). However, the star ratings

from the reviews already cover these topics and give measures that allow ranking the general

satisfaction about these subjects. Hence, these star ratings are more suitable to control for general

satisfaction, compensation and benefits, and work life balance, because topic models only allow

to infer whether a certain topic was discussed or not. Therefore, the approach of the paper is to

define different types of culture based on the literature first and then to check the occurrence of

these different dimensions of culture, instead of using an algorithm to self-generate dimensions of

culture.

4. Empirical Part

4.1 Corporate Culture

Section 4 aims to link corporate culture and divergence from industry-specific culture with firm

performance. Before turning to firm performance, I first present an overview about the

distribution of corporate culture within and between industries, to better understand the

relationship between corporate culture and industry standards. Table 2 shows the correlation of

the different culture dimensions over time. The upper panel presents correlations of firms’ culture

scores with lagged values of the corresponding dimension of culture and they range from 34 –

68%. All reported correlations are significant at the 1% confidence level. This is in line with the

view that organizational culture tends to be stable, but not static (Ott 1989, Hatch 2000). The

4 A further alternative would be to rate a training data set of reviews on their cultural scores and then let a self-

learning algorithm use this training data set to learn what determines these cultural scores, which then can be applied

to the remaining reviews. However, the problem with this approach would be to find objective measures to rate the

reviews of the training data set in the first place.

Page 13: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

12

lower panel analyzes the sector level, which is classified by the GICS5 and shows the continuity

of average scores for culture over time. Most correlations are well above 80%, which suggests

that sector cultures are almost static over time.

TABLE 2. Correlation of Culture over Time

Dimension Innovation Stability Respect Detail Team Outcome Aggressiveness

Panel A: Individual Firms

1st Lag 0.43 0.55 0.38 0.50 0.61 0.61 0.54

2nd

Lag 0.45 0.52 0.40 0.50 0.64 0.55 0.52

3rd

Lag 0.45 0.53 0.34 0.47 0.68 0.43 0.54

Panel B: Sector Averages

1st Lag 0.88 0.93 0.76 0.87 0.91 0.83 0.88

2nd

Lag 0.84 0.90 0.76 0.86 0.86 0.78 0.86

3rd

Lag 0.80 0.90 0.82 0.85 0.82 0.83 0.84 Panel A presents correlations of firms’ culture scores based on the seven dimensions of the OCP, generated by

employee reviews, and the lagged values of the corresponding dimension. Panel B shows correlations of sector

average culture scores and its corresponding lagged values. All reported correlations are significant at the 1% level.

5 See: https://www.msci.com/gics

Figure 2: Mean Culture Scores for the OCP Dimensions Innovation,

Team, and Stability by Sector

Page 14: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

13

Figures 2 and 3 show the mean scores of each dimension of culture by sector. As it can be seen in

different sectors reviewers emphasize words that are related to different cultures: Stability is

particularly highlighted in the utilities sector, teamwork and detail orientation in the information

technology sector, aggressiveness in the energy sector, and the sectors for consumer staples and

discretionary mainly outline outcome orientation and respect for people.

FIGURE 4 AND 5 ABOUT HERE

However, Figures 4 and 5 present boxplots for each dimension of culture and show that

cultures also vary considerably within sectors. To quantify the within and between sector

variation in corporate culture, I conduct analyses of variance (ANOVAs) using firm-level

averages for the culture scores for each dimension, such that ANOVAs do not capture within firm

variation over time. Table 3 presents the mean squares, which are the sum of squares divided by

the degrees of freedom, for ANOVAs. The upper panel defines sector (9 degrees of freedom) as

the group level for ANOVA, while the lower panel analyzes variances within and between sub-

industries (104 degrees of freedom). For both specifications of groups and all dimensions of

culture the null hypothesis that group means are equal can be rejected at the 1% level and the

Figure 3: Mean culture Scores for the OCP Dimensions Outcome

Orientation, Respect for People, Aggressiveness, and Detail

Orientation by Sector

Page 15: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

14

mean squares are always larger between groups than within groups. This is in line with Chatman

and Jehn’s (1994) finding that cultures vary more strongly between groups than within groups.

Finally, I analyze how cultures are distributed and clustered within an industry.

As an example, Figure 6 shows Kernel densities of the culture dimension innovation for the

sectors consumer discretionary, industrials, and materials. Figures AF3 – AF9, in the Appendix,

demonstrate Kernel densities for all dimensions and all sectors. From the figures it can be seen

that cultures tend to cluster around a particular value within an industry, which provides evidence

that indeed industry-specific “standards” of culture exist. Figure 6 demonstrates that the peak of

the distribution tends to be close to the average culture score of a sector. This provides exemplary

evidence that the average industry culture is an appropriate proxy to measure industry standards

of culture.

TABLE 3. Mean Squares between and within Sectors and Sub-Industries from ANOVA

Dimension Innovation Stability Respect Detail Team Outcome Aggressive

Sec

tor Between

groups 6.30 10.956 1.500 8.849 5.305 4.489 6.250

Within

groups .693 .567 .593 .658 .680 .640 .714

Su

b

Ind

ust

ry Between

groups 1.832 1.740 1.063 1.854 1.196 1.459 1.913

Within

groups .52 .497 .492 .532 .650 .513 .527

Total .793 .793 .610 .804 .762 .708 .812

Page 16: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

15

4.2 Corporate Culture and Firm Performance

The previous section shows that cultures vary considerably between and within industries. In this

section I scrutinize these variations to measure its impact on firm performance. Before analyzing

the implications of within industry variation of corporate culture, I consider the general

association between corporate culture and firm performance. This association is estimated by the

following regression equation:

performanceit = 0 + 1*cultureit + 2*culture2it + 3 *ratingit + 4*employeesit + yeart + statei +

sectori + it (1)

The dependent variable is a measure of firm performance, which is either the

firms’ net income or the EBITDA. For both dependent variables one unit corresponds to one

Figure 6: Kernel Density of Innovation for Sectors Discretionary,

Industrials, and Materials. Vertical lines represent sector averages

Page 17: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

16

million US dollar. The variables of interest are reflected in vector cultureit, which refers to the

scores on the seven dimensions of corporate culture measured in Z-scores. I also include second

order polynomials for corporate culture, because the association between culture and firm

performance might not be linear. As the main focus of this study is to analyze, which aspects of

corporate culture a firm should highlight depending on its environment, I want to disentangle the

proposed values from the general satisfaction of employees. For this purpose I include measures

for the star ratings, where employees rate their employer in different dimensions from 1-5. Yet,

these star ratings are naturally highly correlated. An employee, who is satisfied with her career

opportunities, tends to be more satisfied with the senior management and to give her employer a

higher overall rating. To overcome this problem of multicollinearity I conduct a principal

component analysis (PCA), which is reported in Table 4.

TABLE 4. Principal Component Analysis for Star Ratings

Comp1 Comp2 Comp3 Comp4 Comp5

Overall Rating 0.49 -0.05 -0.18 -0.17 -0.83

Work-Life 0.40 0.73 0.49 0.26 0.03

Career 0.47 -0.26 -0.38 0.72 0.23

Comp-Ben 0.40 -0.60 0.65 -0.19 0.17

Leadership 0.47 0.20 -0.40 -0.60 0.47

Eigenvalue 3.78 0.52 0.49 0.12 0.09

The table presents a principal component analysis for employee ratings for different categories

from Glassdoor.

Kaiser (1960) suggests the convention to only include those components that show an

eigenvalue larger than one. However, in the PCA for employee ratings only the first component

has an eigenvalue larger than one. Component 1 shows a considerable association with all

dimensions, but a particularly strong one with the overall rating. Therefore, I label this first

component as PC Rating Overall. To scrutinize further information from employee ratings I also

include Components 2 and 3 in my analysis. Note that my main results do not change when I

follow the convention from Kaiser (1960) and only include Component 1, which has an

eigenvalue larger than one. Component 2 is mainly positively correlated to the ratings for work-

life balance and negatively to compensation and benefits. Therefore, Component 2 is labeled as

PC Work-Life. For Component 3 compensation and benefits is the most important driver, and

Page 18: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

17

therefore labeled for simplification as PC Comp-Ben.6 The vector ratingit contains all of the three

components. Moreover, the number of employees, and fixed effects for year, sector, and the state

of company i’s headquarter7 are included.

TABLE 5 ABOUT HERE

Table 5 presents the results for equation (1) estimated by POLS. For all specifications I

find that PC Rating Overall, which is associated with a high general satisfaction of employees, is

positively correlated with firm performance. This finding is in line with Moniz (2015), who also

exploits reviews from Glassdoor. Similarly, PC Comp-Ben is positively associated with firm

performance. On the other hand, PC Work-Life shows a negative correlation with firm

performance, which is not surprising as it has a strong negative loading of ratings for

compensation and benefits. Columns (1) and (2) analyze the linear relationship between the seven

dimensions of corporate culture and net income or EBITDA, respectively. All dimensions, except

stability, show a positive association with net income, though the effect is only significant for

innovation. The strong positive effect for innovation and the negative effect for stability are in

line with Chatman et al. (2014), who outline the importance of adaptability for firm performance.

Innovation is closely related to adaptability and a strong emphasize on stability might reduce a

firm’s ability to adapt. For EBITDA innovation and aggressiveness have a positive and

significant relationship with firm performance and stability, respect for people, and outcome

orientation a negative, but non-significant association.

In columns (3) and (4) second order polynomials of culture are added. For both measures

of firm performance, six out of seven dimensions of culture show a concave relationship with

firm performance. Bloom, Sadun, and van Reenen (2016) find that management practices serve

as a technology and have a positive and linear effect on firm performance. Columns (3) and (4)

suggest that culture, on the other hand, does not have a linear effect on firm performance, but

6 Note that Glassdoor also provides the opportunity to rate the culture and values of an employer. However, this

dimension is only available for 81% of the firm-year observations. To avoid a reduction in sample size, this

dimension is excluded from my analysis. Yet, most of the variation in the ratings for this dimension is already

captured by PC Rating Overall, because the culture and values ratings show a correlation of 90% with PC Rating

Overall. 7 The states of the companies’ headquarter and the states where most workers are employed overlap for 74% of

observations. I include dummies for the states of firms’ headquarter as these dummies explain slightly more of the

variation in corporate culture. Section 5 discusses the influence of location.

Page 19: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

18

Figure 7: Distribution of Differences to Sector Average for Innovation

rather that there is an “optimal level” of culture. However, the industry-fit hypothesis (Heskett

and Kotter 1992) suggests that the optimal culture depends on industry characteristics, but

equation (1) does not take into account interactions between culture and industry characteristics.

4.3 Corporate Culture and Industry Deviation

So far, the empirical results suggest that cultures differ more strongly between than within

industries and that culture has a concave and sizeable effect on firm performance. Now I turn to

the main question of this paper, which is whether it is beneficial for firms to deviate from the

average culture of their industry.

performanceit = 0 + 1*cultureit + 2*culture2

it + 3*|cultureit -sector_averagejt| + 4*ratingit

+ 5*employeesit + yeart + statei + sectori + it (2)

Page 20: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

19

The coefficient of interest in equation (2) is 3, which measures the absolute difference between

the culture of a firm i from the average culture in its sector j for every dimension of the OCP at

time t. I rely on the absolute difference rather than the squared difference to avoid overestimating

the effect of outliers. Figure 7 shows a histogram for the distribution of the differences (not the

absolute differences) between firms’ corporate culture and their sector’s average score for the

culture dimensions innovation. Figures AF 10 and 11 in the Appendix show distributions of

TABLE 6. Deviation to Average Sector Culture and Firm Performance

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

VARIABLES Net Income EBITDA Net Income EBITDA Net Income EBITDA

Total deviation -278.6*** -522.1*** -510.7*** -984.0***

(81.67) (116.6) (152.4) (317.0)

Innovation deviation -795.1*** -1,284***

(265.9) (380.1)

Stability deviation 777.6*** 371.8

(246.5) (352.3)

Respect deviation -677.7*** -990.5***

(238.0) (340.0)

Outcome deviation -249.5 -61.29

(242.9) (347.2)

Detail deviation -192.1 -386.4

(237.3) (339.1)

Team deviation -680.2** -1,224***

(265.2) (378.9)

Aggressive deviation -174.6 -178.7

(252.2) (360.5)

Avg. wages 5.789 25.16***

(3.712) (7.724)

Constant -1,883 2,551 -1,895 2,340 -2,060 7,519

(2,072) (2,968) (2,078) (2,971) (3,108) (6,467)

Firm Controls YES YES YES YES YES YES

Year FE YES YES YES YES YES YES

Sector FE YES YES YES YES YES YES

State HQ FE YES YES YES YES YES YES

Observations 2,595 2,591 2,595 2,591 559 559

R-squared 0.230 0.377 0.222 0.373 0.541 0.650

Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions are estimated by POLS. The dependent

variable is either net income or EBITDA. Deviation refers to the absolute difference between a firm’s culture and its sector

average for the corresponding dimension of culture. Total deviation corresponds to the sum the deviations for all 7 OCP

dimensions. Firm controls include principal components for star ratings, culture scores, second order polynomials of

culture scores, and the number of employees.

Page 21: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

20

culture scores compared to the sector average for all dimensions and also the combined

differences. As it can be seen, firms’ culture scores are clustered around the average sector score

for corporate culture, though Shapiro-Wilk tests for normality (Shapiro and Wilk 1965) reject the

null-hypothesis that these differences are normally distributed. However, the distribution around

mean-scores of culture supports the hypothesis that within one industry there is one unique

optimal culture rather than multiple equilibria. To disentangle the effect of within industry

deviation from the general effect of culture on performance, I again include control variables for

culture and the second order polynomial of culture.

The regression output for equation (2) is presented in Table 6. In columns (1) and (2) the

effect of deviation from the average industry culture is estimated for each dimension of culture

separately. In column (1) the measure for performance is net income, while in column (2) it is

EBITDA. For 6 of the 7 dimensions, a deviation from the average industry culture is associated

with worse firm performance. This effect is statistically significant for the dimensions innovation,

respect for people, and teamwork. Moreover, these effects are also economically significant. A

one standard deviation increase in culture divergence to the average sector score is associated

with $680k lower yearly net income for teamwork and an over $795k lower yearly net income for

the innovation dimension. However, for the OCP dimension stability a deviation from the

average industry score is associated with better firm performance and this effect is statistically

significant for net income as performance measure.

Columns (3) and (4) report the combined effect of all dimensions. For both outcome

measures the effect is highly significant and results imply that a one standard deviation difference

to the mean industry culture is associated with a $278k decrease in net income and a $522k

decrease in EBITDA. This effect is robust to the exclusion of single sectors. Hence, the results

are not solely driven by one single sector. Columns (5) and (6) include average wage levels as an

additional control variable. Average wages are measured as the total staff expenses divided by the

employment size. COMPUSTAT provides data on staff expenses only for a subset of firms and

therefore the sample size shrinks considerably when including wages as a control variable.

However, for these subsets of firms the effect of overall deviation to the sector average remains

robust.

Table 7 re-estimates equation (2) with the logs of net income and EBITDA as outcome

variables to provide additional interpretations of effect sizes. When these measures of firm

Page 22: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

21

performance are logged, a deviation to the average sector culture is associated with worse firm

performance for all seven dimensions, including stability. This hints that the positive association

of deviation for the stability dimension in Table 6 could be driven by outliers. When combining

all seven dimensions, a one standard deviation increase in deviation to the average industry

culture is associated with 11% lower net income and EBITDA. Again, this effect is robust to the

inclusion of average wages.

TABLE 7 ABOUT HERE

So far, I have defined a business unit by the sector classification from the GICS.

However, the sector classification is potentially too broad to classify business units, because the

business environments and hence the appropriate cultures differ too strongly between various

subunits within a sector. To address this concern, regression output for using the industry-level

and the sub-industry-level classification from the GICS, which are narrower than the sector

classification, is presented in Table 8.

TABLE 8. Deviation to Average Culture for Different Industry Classifications

UNIT Sector Industry Sub-Industry

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

VARIABLES Net Income EBITDA Net Income EBITDA Net Income EBITDA

Total deviation -277.8*** -520.7*** -228.3*** -673.1*** -314.5*** -590.2***

(81.35) (116.1) (74.68) (113.9) (70.72) (116.4)

Controls YES YES YES YES YES YES

Observations 2,595 2,591 2,239 2,235 1,761 1,758

R-squared 0.222 0.373 0.321 0.482 0.443 0.605

Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions are estimated by POLS. The dependent

variable is either net income or EBITDA. Deviation refers to the absolute difference between a firm’s culture and its

sector average for the corresponding dimension of culture. Total deviation corresponds to the sum the deviations for all 7

OCP dimensions. Controls include principal components for star ratings, culture scores, second order polynomials of

culture scores, the number of employees, year fixed effects, state fixed effects, and fixed effects for the corresponding

business unit.

Also with industries and sub-industries as business unites, deviating from the business

units’ average is associated with significantly lower firm performance. For all classifications

Page 23: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

22

business units are excluded when less than 5 firms8 are in the sample, which particularly affects

the sub-industry level. Otherwise, there are too few firms to assume that firms cluster around the

average of their business unit. For the industry and sub-industry classification effect sizes are

very similar to the baseline results that rely on the sector classification and the effects are highly

significant.

4.4 Panel Analysis

The previous results have been entirely estimated by pooled OLS regressions, which take into

account both within and between firm variations in corporate culture. However, the panel

structure of the data also allows for analyzing solely within firm variation in corporate culture to

control for time-invariant omitted variables.

performanceit = Ψi + 1*|cultureit -sector_averagejt| +2*ratingit + 3*employeesit + yeart + it

(3)

Equation (3) estimates the relationship between culture deviation and firm performance with

fixed-effects by including a term Ψi for firm fixed-effects. Hence, coefficients explain how a

within-firm change in the corresponding variable translates into firm performance. Consequently,

the variable of interest 1 can be interpreted as the effect of a firm’s corporate culture diverting

from the sector’s average. Given that sector averages are very stable over time, a change in

corporate culture is highly collinear to a change in deviation to the average sector culture.

Therefore, culture scores are not included as controls in equation (3). Note that including firm

fixed-effects also comes with various caveats. Most importantly, corporate culture tends to be

stable over time and changing corporate culture usually occurs over a longer time span. Yearly

variation in corporate culture might therefore just capture idiosyncratic variation in corporate

8 The minimum number of firms required involves a tradeoff: When the cutoff level is large the sample size shrinks

considerably and when it is low, average scores might not actually reflect the standard culture of the business unit. I

choose a rather small cutoff level of 5, because otherwise too many observations would be lost and when comparing

coefficients it would be not so clear, whether coefficients change because of the different classification or the

reduction in sample size. Hence, Table 8 should be mainly seen as a robustness check rather than a meaningful

comparison of coefficients.

Page 24: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

23

culture and therefore downward bias the effect of corporate culture on firm performance.

Moreover, it is not so clear whether the reviews from a certain year reflects the corporate culture

in that particular year. Employees might report the experience they have made over their whole

employment not just in the recent year. Hence, even if a culture changes, it might take some time

until previous impressions about the firm and its culture vanish. Furthermore, it might also take

some time until an actual change in corporate culture transfers into a change in firm performance.

A culture switch indicates that prevailing structures and codes of interactions are replaced by new

ones. As e.g. Gordon and DiTomaso (1992) and Burt et al. (1994) outline the importance of

shared values, it might take some time until a new culture is established within a firm. Hence, in a

transition period even a change towards a more suitable culture might be associated with worse

firm performance, because the coordination on a new culture creates distortions.9

TABLE 9 ABOUT HERE

Columns (1) and (2) of Table 9 presents results for fixed-effects regressions of equation

(3). As hypothesized above, an increase in total deviation is not associated with a significant

change in firm performance in the corresponding year. An increase in total culture deviation is

associated with a decrease in net income and an increase in EBITDA, but neither of these

relationships is significant. Therefore, columns (3) and (4) consider the lagged effect of an

increase in deviation to the industry difference. For both performance measures, the first lag has

a positive but non-significant effect on firm performance, but the second and third lag have a

negative effect on net income and EBITDA. The effect of the third lag on EBITDA is significant

at the 10% level, while the effect on net income marginally misses the 10% significance level

(p=0.107). These results suggest that in fact culture change is a lengthy process, because it takes

roughly 3 years until a better cultural fit to sector standards transfers into increased firm

performance.

Columns (5) and (6) conduct falsification tests by including leads of total culture

deviation. If in fact culture has a causal effect on firm performance, and not the reverse way, we

should expect to see no association between firm performance and leads of culture deviation. The

results suggest that, if anything, this association is positive. This relationship might stem from

9 The described caveats are less prevalent for POLS regressions as they also consider between-firm variation.

Page 25: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

24

reverse causality: If profits are high, firms choose not to adapt to sector standards. This is in line

with Heskett and Kotter (1992), who find that culture strength was particularly strong at firms,

which previously experienced a growth in firm profits. Note that this bias of reverse causality

would go in the opposite direction compared to the previously reported relationship of decreased

firm performance due to an increase in culture deviation. Hence, the results from pooled OLS

regressions are potentially downward biased.

4.5 Corporate Culture and Executives

When a deviation from the average culture of the business unit is associated with worse firm

performance, why do firms within one industry offer different corporate cultures? A potential

explanation for varying corporate cultures could be CEOs and executives, who impose

suboptimal cultures.

This would be in line with Bloom and van Reenen (2010) as the results from part 4.3 and 4.4

suggest that offering an appropriate culture might be a further channel of how management

practices affect firm performance. Therefore, I analyze how CEO and executives characteristics

are associated with corporate culture. For this, I match the culture data and firm characteristics

with the Compustat Executive Compensation (Execucomp) database, which contains information

regarding the salary and bonus of executives but also their age and the duration a CEO has been

in office. As the compensation of executives is also strongly driven by firm performance I focus

TABLE 10. Corporate Culture and Executives

(1) (2) (3) (4) (5) (6) (7)

VARIABLES Innovation Stability Respect Detail Team Outcome Aggressive

Executives Age -0.0066 0.027*** -0.0071 7.03e-05 -0.0191*** -0.0056 -0.0012

(0.0068) (0.0060) (0.0064) (0.0074) (0.0069) (0.0058) (0.0072)

CEO Tenure -0.0049 -0.011*** 4.41e-05 0.0020 -0.00010 0.0018 -0.0018

(0.0035) (0.0033) (0.0033) (0.0038) (0.0036) (0.0030) (0.0037)

Firm Controls YES YES YES YES YES YES YES

Sector FE YES YES YES YES YES YES YES

Year FE YES YES YES YES YES YES YES

Observations 2,988 2,988 2,988 2,988 2,988 2,988 2,988

Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions are estimated by POLS. The

dependent variables refer to single dimensions of the OCP. Firm controls include principal components for star

ratings, culture scores, second order polynomials of culture scores, and the number of employees.

Page 26: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

25

on the average age of executives and the tenure of CEOs. Table 10 presents the association of

these two characteristics with the seven culture dimensions. In all columns I control for year- and

sector fixed effects, as well as for number of employees and rating scores from Glassdoor. Older

executives are associated with less teamwork but more stability. A long CEO tenure is negatively

correlated with stability. Executives’ age and CEO tenure are both negatively associated with

innovation, though this effect is not significant. The negative association of both variables with

innovation and the positive correlation of executives’ age and stability is in line with previous

findings showing that older CEOs tend to prefer less risky firm policies (Serfling, 2014), though

the negative association of CEO tenure and stability point in a different direction.

Table 11 analyzes how CEO characteristics along with other firm characteristics are

associated with deviation to the average sector culture. Hence, Table 11 reports regression results

TABLE 11. Culture Deviation and Firm Characteristics

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

VARIABLES Total deviation Total deviation Total deviation Total deviation

Number Competitors -0.0145*** -0.0148*** -0.0144*** -0.0151***

(0.00396) (0.00391) (0.00396) (0.00392)

Employees -0.00119*** -0.00122*** -0.00119*** -0.00116***

(0.000188) (0.000187) (0.000188) (0.000188)

PC Rating Overall -0.0399*** -0.0387** -0.0412*** -0.0417***

(0.0152) (0.0151) (0.0152) (0.0151)

PC Work-Life -0.0275 -0.0510 -0.0338 -0.0326

(0.0452) (0.0444) (0.0456) (0.0446)

PC Comp-Ben -0.0129 -0.0138 -0.0194 -0.00684

(0.0404) (0.0398) (0.0409) (0.0402)

CEO Tenure 0.00878** 0.00773**

(0.00370) (0.00382)

Executives‘ Age 0.0102 0.00791

(0.00694) (0.00727)

CEO Switch -0.123**

(0.0510)

Constant 3.039*** 2.460*** 2.573*** 3.049***

(0.540) (0.680) (0.689) (0.537)

Sector FE YES YES YES YES

State HQ FE YES YES YES YES

Year FE YES YES YES YES

Observations 1,990 2,021 1,990 2,026

R-squared 0.699 0.698 0.699 0.699 Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions are estimated by POLS. The dependent

variable is the combined deviation of all culture dimensions to the average sector cultures. Number competitors refers

to the number of firms in the sample that are in the same subindustry. CEO Switch is a dummy variable that is equal to

one if the firm had a CEO switch in the previous 5 years.

Page 27: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

26

with the total deviation to the sector culture as outcome variable. First, those firms that differ

strongly to their sector’s culture have a significantly lower employment size. Table 11 also

analyzes the effect of market competition on the deviation to the average industry culture.

Competition is measured by the number of firms in the same sub-industry in this sample10

.

Results indicate a negative and significant correlation between total deviation and market

competition. Similar to these results, Bloom and Van Reenen (2007) find that firm size and

product market competition are both positively correlated with management practices. Moreover,

the total deviation is associated with lower ratings on the Glassdoor star rating as all principal

components show a negative correlation with total deviation and this effect is highly significant

for PC Rating Overall, the component, which mainly captures the overall satisfaction with an

employer. Again, this is in line with Bloom, Kretschmer, and Van Reenen (2009), who find that

in well-managed firms the work-life balance tends to be better for employees. The different

specifications in Table 11 include different measures of CEO. In specification (1) CEO tenure

shows a positive and significant correlation with the total deviation to the average culture. This

association also holds when including the average age of executives as an independent variable in

specification (3). Similarly, Miller (1991) finds that long-tenured CEOs are less able to adapt to

their firm’s environment. The average executives’ age shows a positive but not significant

correlation with total deviation in specifications (2) and (3). Specification (4) includes an

indicator variable that is equal to one, if the firm experienced a CEO switch in the last 5 years.

Firms that recently experienced a switch in CEO show a lower deviation to the average sector

culture and this correlation is highly significant.

All of these results are related to Bloom and Van Reenen (2007) and lead to similar

implications: According to Bloom and Van Reenen (2007) managers can influence firm

performance by applying good management practices. My results suggest that managers, and

particularly CEOs, also have an influence on corporate culture. Therefore, a further channel on

how managers can influence firm performance is by imposing a culture that is suitable for their

firm’s environment. The above shown association between executives’ characteristic and

10 The association between market competition and industry interaction in corporate culture could be examined in

more detail in future research by applying more precise measures of market competition as the Hirschmann-

Herfindahl-Index (Hirschmann 1964)

Page 28: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

27

corporate culture also adds to the debate whether corporate culture and organizational change are

mainly imposed bottom-up or top-down (Rashid, Zabid, and Rahman 2004). Similar to Berson

and Oreg (2008) my results outline the important role of CEOs in shaping and changing corporate

culture.

5. Regional Culture

5.1 Regional Characteristics and Corporate Culture

Besides industry characteristics, corporate cultures could also vary due to differences in regional

culture. Hofstede (1983), who analyzes differences in corporate culture between countries, argues

that a corporate culture is always embedded in its external environment, including the dominant

cultural traits in the region of a firm and its workers. Though this study is deliberately limited to

the United States to abstract from variation between countries, different regional cultures within

the United States might still affect corporate culture. At first I will show that regional cultures do

matter for corporate culture, but that this association is small compared to industry

characteristics. Moreover, I provide evidence that the association between deviation from the

average industry culture and firm performance is not driven by deviation from the regional

cultures.

Table 12 compares the influence of the firms’ location on their culture with the effect of

sector and sub-industry characteristics. To do so, I conduct regressions of each dimension of

corporate culture, where only dummies for the location of the firm, its sector or its sub-industry

are included. For the location of the firm I include both dummies for state of headquarter and

secondly a dummy for the state, where according to the employee reviews most workers are

employed. However, for 74% of the firms these states are the same. As a comparison, I also

conduct regressions that only include dummies for the sector or sub-industry. Table 12 reports

how much of the variation in corporate culture can be explained by these sets of dummies

respectively. For most dimensions the dummies for industries explain considerably more of the

variation in corporate culture. Moreover, it seems that the headquarter’s location of a firm is more

important for corporate culture than the location of employees, which is in line with Schein

Page 29: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

28

(1992), who outlines the importance of leaders for the formation, conservation, and alteration in

corporate culture and argues that culture is mainly imposed top-down.

To find some further evidence about the effect of regional culture on corporate culture I

match the data on corporate culture with survey data from the World Value Survey Wave 6

(2010-2014). The World Value Survey (WVS) is a global research project, which examines

various beliefs and values about, among other things: social preferences, politics, economics, and

religion. 2232 people from the United States completed Wave 6 of the WVS and the WVS also

keeps record about the state in which the survey was conducted. Since for some states less than

10 participants completed the survey, I analyze regional values at the divisional level and follow

the classification of divisions from the U.S. Census Bureau11

, which separates the United States

into 9 divisions. As a measure for regional culture I rely on Schwartz’s Value Inventory

(Schwartz 1992), which is included in the WVS and measures values regarding: creativity,

money, security, hedonism, benevolence, success, tradition, risk-taking, and the environment. At

first, I conduct a principal component analysis of these values, because they are strongly

correlated among each other. Table 13 reports these principal components. I only consider those

eigenvalues that are larger than one (Kaiser 1960). The first component is mainly associated with

values regarding creativity, hedonism, and the environment (label: PC Creativity), while the

second components relates to security, proper behavior, and traditions (label: PC Security).

11 See: https://www.census.gov/geo/reference/gtc/gtc_census_divreg.html

TABLE 12. Corporate Culture and Indicators for Industry and Region

Dimension Innovation Stability Respect Detail Team Outcome Aggressive

HQ-State R2 .0725 .0805 .0464 .0957 .0873 .0667 .0904

Adj. R2 .0494 .0576 .0227 .0732 .0646 .0435 .0677

Work-State R2 .0585 .0805 .0432 .0769 .0909 .0537 .0660

Adj. R2 .0411 .0634 .0254 .0597 .0740 .0361 .0486

Sector R2 .1005 .1865 .0247 .1679 .0896 .0903 .1143

Adj. R2 .0959 .1825 .0198 .1638 .0851 .0857 .1099

Subindustry R2 .3399 .3588 .1322 .3551 .2511 .2626 .3502

Adj. R2 .2997 .3198 .0794 .3158 .2055 .2177 .3107 For each pair of explanatory variables (Dummies for HQ-State, Work-State, Sector, Sub-Industry and dimension

of culture) an OLS regression of that dimension of culture on the corresponding explanatory variables was

conducted. For HQ-State dummy variables for the state the firm’s headquarters are located is included. Work-

State includes dummies for the state where most workers are employed. Sector and Sub-Industry include dummies

for sector or sub-industry according to the GICS classification respectively. Each cell reports the R2 or adjusted

R2 of these regressions.

Page 30: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

29

TABLE 13 ABOUT HERE

On this basis, I regress all culture dimensions on the 2 principal components. Note that the

average values are matched to the respective division based on the headquarter location of each

firm. However, since for most firms the headquarter location also corresponds to the state, where

most workers are employed, results only vary sparsely when the matching is conducted based on

the location of workers. For both cases the explanatory power of these measures of regional

culture is limited. In regressions without further controls R2s range from only 0.1% to 2.7%.

TABLE 14. Corporate Culture and Regional Culture

(1) (2) (3) (4) (5) (6) (7)

VARIABLES Innovation Stability Respect Detail Outcome Team Aggressive

PC Creativity 0.0078 -0.018* 0.035*** 0.036*** 0.0017 0.036*** 0.031***

(0.010) (0.010) (0.011) (0.011) (0.011) (0.011) (0.011)

PC Security -0.040** 0.016 -0.022 -0.021 0.030* 0.0050 0.010

(0.017) (0.017) (0.019) (0.018) (0.018) (0.018) (0.018)

Employees 0.00013 0.00054*** 0.00047*** -8.84e-05 -7.63e-05 -0.0003** -0.00042***

(0.00014) (0.00014) (0.00016) (0.00015) (0.00015) (0.00015) (0.00015)

Sector FE YES YES YES YES YES YES YES

Observations 1,588 1,588 1,588 1,588 1,588 1,588 1,588

R-squared 0.135 0.234 0.041 0.174 0.091 0.103 0.127

Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions are estimated by POLS. The dependent

variables refer to single dimensions of the OCP. PC Creativity and PC Security refer to principal components that are

based on the Schwartz’s Value Inventory from the World Value Survey Wave 6.

Table 14 reports results for OLS regressions of all dimensions of culture on regional

values with further controls for the industry and number of employees. For all dimensions at least

one of the two principal components is significantly associated with corporate culture. PC

Creativity, which is mainly associated with creativity, hedonism, and the environment, is

negatively correlated with stability and positively correlated with respect for people, detail

orientation, team orientation, and aggressiveness. PC Security, which is mainly related to values

regarding security, proper behavior, and traditions, is negatively associated with innovation and

positively correlated with outcome orientation. On the whole, these associations seem adequate

and suggest that also within countries, regional cultures do matter for corporate culture. However,

these results should be taken with a grain of salt as the above shown results only stem from

variation in regional culture scores between 9 U.S. divisions, and therefore the association of

regional culture and corporate culture requires more in depth research.

Page 31: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

30

5.2 Regional Deviation and Firm Performance

Given the weak but existing evidence for an association between regional culture and corporate

culture I now check whether my main results regarding within industry variation of corporate

culture is actually driven by regional culture. The negative association between deviation from

industry-culture and firm performance might confound with regional distribution of firms. As

business units are often clustered in particular areas (Feser and Bergman, 2000), the detrimental

effect of deviating from the average culture of the respective field of business, might actually

stem from divergence from regional culture. To control for this potential confound, I re-estimate

equation (2), but this time I consider the deviation of firm’s culture to the average culture of all

firms that have its headquarter in the same state. When for a given state in a given year less than

10 firms are available, these observations are dropped. Table 15 reports the overall effect of this

deviation (regional culture deviation) on firm performance. Columns (1) and (2) show these

effects without controlling for sector deviation. The association between deviation to regional

culture and firm performance is negative but not significant for both measures of firm

performance. When I include total deviation, which still corresponds to the deviation from the

average sector culture, as an independent variable, the effect of regional culture deviation

remains insignificant. On the other hand, the effect of deviation to the average sector culture and

firm performance shrinks slightly compared to the previous results, but remains negative and

TABLE 15. Regional Deviation and Firm Performance

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

VARIABLES Net Income EBITDA Net Income EBITDA

Region Culture Deviation -36.99 -229.5 76.22 -73.02

(93.80) (147.5) (109.1) (171.6)

Total Deviation -197.7** -273.3*

(97.68) (153.6)

Controls YES YES YES YES

Observations 2,595 2,591 2,595 2,591

R-squared 0.264 0.430 0.266 0.431 Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions are estimated by POLS.

The dependent variable is either net income or EBITDA. Regional Culture Deviation refers to the

combined deviation of firm’s culture to the average culture of the state where its headquarter is located.

Total deviation corresponds to the culture deviation from the industry average. Controls include principal

components for star ratings, culture scores, second order polynomials of culture scores, the number of

employees, year fixed effects, state fixed effects, and sector fixed effects.

Page 32: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

31

significant, when controlling for deviation to the average regional culture. This provides evidence

that in fact the characteristics of the industry rather than regional culture matters mostly for

corporate cultural fit.

6. Robustness Checks

The main finding of this paper suggests that a deviation from the average industry culture is

associated with worse firm performance. This section provides evidence that this finding is robust

to the amount of reviews required for a firm-year observation, the imputation of culture scores for

missing values, and taking into account negations in employee reviews.

One concern regarding the interpretation of the results is that they could be driven by the

amount of reviews that generate culture scores for a given firm-year observation. In the empirical

section I include all observations that are generated by at least 30 reviews. Yet, a potential

explanation for the findings could be that observations, which are generated by only a few

employee reviews on Glassdoor, naturally show a higher variance and therefore also vary on

average more strongly from the industry average. On the other hand, the amount of reviews per

year is not exogenous. Consequently, dropping observations that originate from only few reviews

might lead to a sample selection bias. The initial cutoff level of 30 reviews was chosen to address

this tradeoff. It excludes those observations where inferences on corporate culture are hardly

possible due to limited reviews, but includes sufficiently many observations to not only analyze a

small and highly selective sample. Columns (1) and (2) of Table 16 re-estimate equation (2) for

different cutoff levels to analyze, if the main results are driven by the cutoff level. Every cell in

Table 16 corresponds to one individual regression, but only reports the coefficient for the variable

of interest, total deviation. Yet, similar to Table 6 it always controls for the employment size,

employee ratings, culture scores for the seven dimensions of the OCP and their second order

polynomials, sector fixed effects, year fixed effects, and dummies for the state of the companies’

headquarter location. It analyzes the effect of deviating from the average sector culture, when

different cutoffs levels are applied. Specifically, it tests the robustness of the results for the cutoff

levels 0, 30, 50, 100, and 150. For all cutoff levels total deviation has a negative and highly

significant effect on net income and EBITDA. Effect sizes rise with higher cutoff levels and for a

Page 33: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

32

cutoff level of 150 the effect size increases with a factor of roughly 2.6 for net income and 2 for

EBITDA compared to the baseline results (cutoff level =30). This suggests that baseline results

are potentially downward biased, because observations generated by few reviews contain more

idiosyncratic variance around their true cultural scores. Columns (3) and (4) address the concern

of sample selection bias, as the amount of reviews generated might be endogenous to firm

performance. Therefore, I impute values for the seven cultural dimensions, when a firm-year

observation lies below the corresponding cutoff level. The imputed values are generated by a

function of the average values for the seven culture dimensions. Effect sizes are smaller

compared to baseline results without imputation. However, for all cutoff levels and both

performance measures total deviation is still associated with worse firm performance. These

effects are highly significant for all cutoff levels when the performance measure is EBITDA and

for cutoff levels below 100 for net income.

Finally, columns (5) and (6) address the concern of negative statements. The word count

method that is scrutinized in the baseline regressions does not take into account whether the

words from the master texts are used in positive or negative contexts. For instance, when many

employees outline that the company is “not innovative“ or “not very modern” the text similarity

algorithm would give its culture a high score on innovation. To address this concern, I analyze

whether the word related to culture is stated in combination with one of the following negation

terms12

: “not, don't, no, none, nobody, nothing, neither, nowhere, never, hardly, barely, scarcely,

doesn't, isn't, wasn't, shouldn't, wouldn't, couldn't, won't”13

. To do so I analyze every sentence

and subordinate clause independently and when a sentence contains one of the above mentioned

negation terms, words related to culture are weighted negatively for the text similarity score.

Columns (5) and (6) present results for the coefficient total deviation when estimating

equation (2) with this adjustment for negations. Again, total deviation is always associated with

lower firm performance and this effect is significant for all cutoff levels. Effect sizes are similar

to the dataset that does not take into account negations, which provides evidence that negative

statements are not an issue when interpreting the baseline results.

12 Hogenboom et al. (2011) discuss different approaches for dealing with negations.

13 These words are not case-sensitive, meaning that the word “wont” would equally count as a negation.

Page 34: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

33

TABLE 16. Robustness Checks: Coefficients of Total Deviation

Basic Imputation Negative Statements

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

Net Income EBITDA Net Income EBITDA Net Income EBITDA

cutoff=0 -161,83*** -420,97*** -162,42*** -422,12*** -144,88*** -397,93***

(40,29) (60,15) (40,36) (60,26) (34,42) (51,50)

n=4930 n=4930 n=4930 n=4930 n=4930 n=4930

cutoff=30 -279,63*** -523,88*** -125,42*** -271,22*** -283,23*** -743,42***

(81,91) (116,92) (30,68) (45,84) (103,85) (147,08)

n=2663 n=2663 n=4930 n=4930 n=2663 n=2663

cutoff=50 -376,06*** -603,86*** -111,39*** -248,08*** -386,66*** -761,44***

(105,20) (147,04) (30,51) (45,59) (103,07) (182,09)

n=1973 n=1973 n=4930 n=4930 n=1973 n=1973

cutoff=100 -588,28*** -816,65*** -61,04** -174,58*** -395,43*** -916,78***

(130,91) (233,01) (30,22) (45,21) (151,16) (272,08)

n=1158 n=1158 n=4930 n=4930 n=1158 n=1158

cutoff=150 -742,10*** -999,00*** -28,35 -124,72*** -390,23* -720,55*

(191,45) (341,71) (29,78) (44,64) (217,68) (384,93)

n=761 n=761 n=4930 n=4930 n=761 n=761 Significance: *** p<0.01, ** p<0.05, * p<0.1. Standard errors in parentheses. All regressions run by POLS. Each

cell represents a single regression and reports coefficients for total sector deviation. All regressions control for

employment size, employee ratings, culture scores, second order polynomials on culture scores, year fixed effects,

state HQ fixed effects, sector fixed effects. Basic re-estimates Table 6 for different cutoff levels, where observations

that were generated by less reviews than the cutoff level are dropped. Imputation, imputes values for observations

that lie below this cutoff. Negative Statements does not impute values for dropped observations, but takes into

account when a word related to culture was used in a negative context.

7. Conclusion

In a survey conducted by Strategy& (formerly Booz & Company) 84% of participants agreed that

corporate culture is critical to business success. However, 45% of respondents stated that their

culture is not effectively managed (Aguirre, von Post and Alpern 2013). Therefore, the recent

interest of both economist and organizational psychologists in linking corporate culture with firm

performance is not surprising. Work on this includes culture strength (Sørensen 2002),

trustworthiness (Guiso, Sapienza, and Zingales 2015) and adaptability (Chatman et al. 2014).

This paper builds on this literature, outlining that the effect of culture on performance might

differ between industries, and analyzes the effect of industry characteristics and standards in

more detail. Previous theories and findings contradict in their implications regarding how

Page 35: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

34

divergence from the average industry culture is associated with firm performance. This paper

shows that a deviation from the average industry is associated with reduced firm performance.

This effect is statistically and economically significant. I present evidence that this effect

is robust to wage levels, the general satisfaction of employees, regional differences, and the direct

effect of culture itself. As a measure of corporate culture I rely on modern web scraping and text

mining approaches and scrutinize employee reviews from the employee review site,

Glassdoor.com. The distribution of ratings and the 7 dimensions of culture indicate that these

publicly available information are in fact informative about corporate culture. Moreover, I show

that my results are not sensitive to negations in these online reviews or the decision rule about

how many reviews are required to include a firm in the sample. Relying on online reviews instead

of surveys has two major advantages: First, it allows gathering a large sample of firms and

therefore to analyze within industry variation. Second, it includes a more diverse set of

employees, as surveys regarding corporate culture are usually conducted with executives.

However, the approach of scrutinizing online reviews could be extended in future research by

extracting job titles of employees and comparing characteristics of reviewers with characteristics

of the general workforce.

Moreover, I find that mainly executives’ characteristics explain differences in culture,

while regional differences only play a minor role. This provides evidence that corporate culture is

imposed top-down and not developed bottom-up. This relationship could be examined in more

detail in future research as well. For instance, one could analyze how exogenous CEO changes,

by death or retirement, affect corporate culture.

Page 36: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

35

Figure 4: Boxplots for the Culture Dimensions Innovation, Outcome Orientation, Stability, and Respect for

People by Sector

Page 37: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

36

Figure 5: Boxplots for the Culture Dimensions Detail

Orientation, Teamwork, and Aggressiveness by Sector

Page 38: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

37

TABLE 1. Characteristics of Glassdoor Users based on Quantcast Characteristic Composition Index

Place

United States 87%

Rest of World 13%

Gender

Male 45% 90

Female 55% 110

Age

<18 12% 67

18-24 20% 142

25-34 24% 150

35-44 20% 103

45-54 16% 92

55-64 7% 70

65+ 1% 26

Household Income

$0-50k 45% 88

$50-100k 30% 100

$100-150k 14% 119

$150k+ 11% 147

Education Level

No College 26% 55

College 52% 129

Grad School

22% 159

Ethnicity

Caucasian 64% 83

African American 13% 157

Asian 9% 216

Hispanics 12% 125

Other 2% 116

The Table shows characteristics of Glassdoor users that were estimated by Quantcast. Index shows, how these

demographics correspond to the general internet population, where an index of 100 represents an exact

representation.

Page 39: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

38

TABLE 5: Corporate Culture and Firm Performance

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

VARIABLES Net Income EBITDA Net Income EBITDA

PC Rating Overall 511.0*** 921.6*** 497.2*** 901.7***

(84.32) (123.8) (84.70) (124.3)

PC Work-Life -685.5*** -2,079*** -647.5*** -2,036***

(206.6) (303.3) (207.2) (303.8)

PC Comp-Ben 261.8 1,031*** 198.7 933.4***

(210.4) (308.9) (211.1) (309.7)

Innovation 321.1** 386.4* 360.9** 472.5*

(147.3) (216.3) (179.1) (262.7)

Stability -2.676 -229.8 353.8* 230.7

(140.7) (206.9) (182.7) (268.0)

Respect 53.82 -227.1 159.3 -72.99

(143.4) (210.6) (172.0) (252.2)

Outcome 17.32 -185.1 -69.78 -282.3

(163.4) (240.4) (203.2) (298.4)

Detail 26.77 45.86 249.7 313.1

(130.7) (191.8) (191.2) (280.3)

Team 38.83 188.5 259.5 675.6***

(131.2) (192.6) (164.9) (241.8)

Aggressive 214.0 417.0** 153.2 319.7

(134.1) (196.8) (174.2) (255.5)

Innovation^2 -35.99 -80.36

(78.53) (115.1)

Stability^2 -238.9*** -305.4***

(74.11) (108.7)

Respect^2 -68.88 -107.8

(79.33) (116.3)

Outcome^2 58.17 41.02

(88.12) (129.2)

Detail^2 -94.92* -117.8

(54.56) (80.00)

Team^2 -135.7** -293.3***

(56.59) (82.97)

Aggressive^2 -2.579 -5.023

(59.08) (86.64)

Employees 8.795*** 21.33*** 8.600*** 21.06***

(0.840) (1.233) (0.840) (1.232)

Sector FE YES YES YES YES

Year FE YES YES YES YES

State HQ FE YES YES YES YES

Observations 2,366 2,666 2,366 2,666

R-squared 0.170 0.345 0.180 0.355 Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions are estimated by POLS. The

dependent variable is either firms’ net income or EBITDA. PC Rating Overall, PC Work-Life, PC Comp-Ben

correspond to principal components regarding the employee ratings from Glassdoor. Innovation, Stability,

Respect, Outcome, Detail, Team, Aggressive refer to Z-scores for the corresponding dimension of the OCP and

were derived from Glassdoor reviews. State HQ refers to the state in which the companies headquarter is located.

Page 40: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

39

TABLE 7. Deviation to Average Sector Culture and Logs of Firm Performance

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

VARIABLES Log

Net Income

Log

EBITDA

Log

Net Income

Log

EBITDA

Log

Net Income

Log

EBITDA

Total deviation -0.115*** -0.113*** -0.149*** -0.124***

(0.0210) (0.0175) (0.0440) (0.0354)

Innovation deviation -0.229*** -0.172***

(0.0700) (0.0575)

Stability deviation -0.0167 -0.0421

(0.0665) (0.0554)

Respect deviation -0.124* -0.157***

(0.0642) (0.0523)

Outcome deviation -0.0840 -0.0457

(0.0655) (0.0536)

Detail deviation -0.112* -0.0965*

(0.0637) (0.0524)

Team deviation -0.151** -0.175***

(0.0688) (0.0577)

Aggressive deviation -0.0823 -0.101*

(0.0665) (0.0558)

Avg. Wages 0.00391*** 0.00514***

(0.00102) (0.000871)

Firm Controls YES YES YES YES YES YES

Year FE YES YES YES YES YES YES

Sector FE YES YES YES YES YES YES

State HQ FE YES YES YES YES YES YES

Observations 2,366 2,666 2,366 2,666 498 558

R-squared 0.534 0.565 0.533 0.564 0.791 0.818

Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions are estimated by POLS. The dependent

variable is either the log of net income or the log of EBITDA. Deviation refers to the absolute difference between a firm’s

culture and its sector average for the corresponding dimension of culture. Total deviation corresponds to the sum the

deviations for all 7 OCP dimensions. Firm controls include principal components for star ratings, culture scores, second

order polynomials of culture scores, and the number of employees.

Page 41: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

40

TABLE 9. Fixed Effects Regression. Deviation to Average Sector Culture and Firm Performance

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

VARIABLES Net Income EBITDA Net Income EBITDA Net Income EBITDA

Total deviation -88.04 25.37

(81.06) (73.51)

L. Total deviation 67.53 8.881

(53.15) (59.17)

L2. Total deviation -27.50 -46.76

(49.45) (55.03)

L3. Total deviation -71.09 -85.48*

(44.03) (48.99)

F. Total deviation 229.4* 61.25

(119.2) (95.49)

F2. Total deviation 79.12 -83.48

(131.5) (105.3)

F3. Total deviation 23.78 148.2

(150.1) (120.2)

Firm Controls YES YES YES YES YES YES

Firm FE YES YES YES YES YES YES

Year FE YES YES YES YES YES YES

Observations 2,988 2,980 1,214 1,210 1,195 1,194

R-squared 0.068 0.104 0.059 0.117 0.085 0.128

Number of id 668 666 376 376 366 366 Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All regressions are estimated by Fixed Effects

Regression. The dependent variable is either net income or EBITDA. Deviation refers to the absolute difference

between a firm’s culture and its sector average for the corresponding dimension of culture. Total deviation

corresponds to the sum of deviations to average sector scores for all 7 OCP dimensions. L.,L2., L3. refer to lagged

values and F.,F2., F3. refer to future values (leads). Firm controls include principal components for star ratings, and

the number of employees.

Page 42: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

41

TABLE 13: Principal Component Analysis: Schwartz's Value Inventory from WVS

Variable Comp1 Comp2 Comp3 Comp4 Comp5 Comp6 Comp7

Creativity 0.418 0.110 0.391 0.028 0.154 -0.401 -0.331

Wealth 0.319 0.193 -0.619 -0.324 -0.318 -0.119 0.156

Secure -0.090 0.586 -0.153 0.369 -0.194 -0.531 0.070

Hedonism 0.438 0.168 -0.069 -0.158 0.017 0.424 0.270

Success 0.362 0.354 0.163 -0.302 0.016 0.153 -0.493

Risk 0.331 -0.063 -0.198 0.787 -0.114 0.367 -0.217

Proper behavior -0.213 0.496 -0.189 0.055 0.742 0.216 0.063

Tradition -0.225 0.445 0.497 0.018 -0.473 0.351 0.215

Environment 0.432 -0.069 0.296 0.136 0.222 -0.187 0.669

Eigenvalue 4.637 2.347 0.872 0.691 0.313 0.102 0.034

Page 43: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

42

References

Abdul Rashid, Zabid, Murali Sambasivan, and Azmawani Abdul Rahman. 2004. "The

influence of organizational culture on attitudes toward organizational change." Leadership &

organization development Journal. 25(2): 161-179.

Aguirre, DeAnne, Rutger von Post, and Micah Alpern. 2013. "Culture's Role in Enabling

Organizational Change Survey Ties Transformation Success to Deft Handling of Cultures

Issues." Berlin: Booz & Company

Berson, Yair, Shaul Oreg, and Taly Dvir. 2008. "CEO values, organizational culture and firm

outcomes." Journal of Organizational Behavior 29(5): 615-633.

Bertrand, Marianne, and Antoinette Schoar. 2003. "Managing with style: The effect of

managers on firm policies." The Quarterly Journal of Economics, 118(4): 1169-1208.

Bertrand, Marianne and Sendhil Mullainathan. 2004. "Are Emily and Greg more employable

than Lakisha and Jamal? A field experiment on labor market discrimination." American

economic review, 94(4): 991-1013.

Blei, David M., Andrew Y. Ng, and Michael I. Jordan. 2003. “Latent dirichlet allocation.”

Journal of machine Learning research, 3(Jan): 993–1022.

Bloom, Nicholas, Raffaella Sadun, and John van Reenen. 2016. “Management as a

Technology?” National Bureau of Economic Research.

Bloom, Nicholas and John van Reenen. 2010. “Why do management practices differ across

firms and countries?” The Journal of Economic Perspectives, 24(1): 203–24.

Bloom, Nick, Tobias Kretschmer, and John Van Reenan. 2009. "Work-life balance,

management practices and productivity." International differences in the business practices

and productivity of firms. University of Chicago Press, 15-54.

Bloom, Nicholas, and John Van Reenen. 2007. "Measuring and explaining management

practices across firms and countries." The Quarterly Journal of Economics, 122(4): 1351-

1408.

Burt, Ronald S., Shaul M. Gabbay, Gerhard Holt, and Peter Moran. 1994. “Contingent

organization as a network theory. The culture-performance contingency function.” Acta

Sociologica, 37(4): 345–70.

Chatman, Jennifer A. 1991. “Matching People and Organizations: Selection and Socialization

in Public Accounting Firms.” Administrative Science Quarterly, 36.

Chatman, Jennifer A., David F. Caldwell, Charles A. O'Reilly, and Bernadette Doerr. 2014.

“Parsing organizational culture. How the norm for adaptability influences the relationship

between culture consensus and financial performance in high‐technology firms.” Journal of

Organizational Behavior, 35(6): 785–808.

Chatman, Jennifer A. and Karen A. Jehn. 1994. “Assessing the relationship between industry

characteristics and organizational culture. How different can you be?” Academy of

management journal, 37(3): 522–53.

Crémer, Jacques. 1993. “Corporate culture and shared knowledge.” Industrial and corporate

change, 2(3): 351–86.

Page 44: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

43

Compustat Industrial [Annual Data]. (2008-2016). Available: Standard & Poor's/Compustat.

Retrieved from Wharton Research Data Service.

Dessein, Wouter and Andrea Prat. 2017. "Organizational Capital, Corporate Leadership, and

Firm Dynamics."

Feser, Edward J., and Edward M. Bergman. 2000. "National industry cluster templates: a

framework for applied regional cluster analysis." Regional studies, 34(1):1-19.

Ghose, Anindya, and Panagiotis G. Ipeirotis. 2011. "Estimating the helpfulness and economic

impact of product reviews: Mining text and reviewer characteristics." IEEE Transactions on

Knowledge and Data Engineerin,g 23(10): 1498-1512.

Gordon, George G. and Nancy DiTomaso. 1992. “Predicting corporate performance from

organizational culture.” Journal of management studies, 29(6): 783–98.

Guiso, Luigi, Paola Sapienza, and Luigi Zingales. 2008. “Social capital as good culture.”

Journal of the European Economic Association, 6(2-3): 295–320.

Guiso, Luigi, Paola Sapienza, and Luigi Zingales. 2015. “The value of corporate culture.”

Journal of Financial Economics, 117(1): 60–76.

Hansen, Stephen and Michael McMahon. 2016. “Shocking language. Understanding the

macroeconomic effects of central bank communication.” Journal of International Economics,

99: S114-S133.

Hartnell, Chad A., Amy Y. Ou, and Angelo Kinicki. 2011. “Organizational culture and

organizational effectiveness: a meta-analytic investigation of the competing values

framework's theoretical suppositions.” The Journal of applied psychology, 96(4): 677–94.

Heskett, James L. and John P. Kotter. 1992. Corporate culture and performance. New

York: Free Press.

Hatch, Mary Jo. 2000. The cultural dynamics of organizing and change. Handbook of

organizational culture and climate. 245-260.

Hirschman, Albert O. 1964. "The paternity of an index." The American Economic Review,

54(5): 761-762.

Hofstede, Geert. 1983. “The cultural relativity of organizational practices and theories.” Journal

of international business studies, 14(2): 75–89.

Hogenboom, Alexander, Paul van Iterson, Bas Heerschop, Flavius Frasincar, and Uzay

Kaymak. 2011. “Determining negation scope and strength in sentiment analysis.”

Hyponym. 2018. In Oxford Dictionaries Online. Retrieved February 18, 2018, from

https://en.oxforddictionaries.com/definition/hyponym

Kaiser, Henry F. 1960. “The application of electronic computers to factor analysis.”

Educational and psychological measurement, 20(1): 141–51.

Kosfeld, Michael and Ferdinand A. von Siemens. 2011. “Competition, cooperation, and

corporate culture.” The RAND Journal of Economics, 42(1): 23–43.

Kreps, David. 1990. Perspectives on Positive Political Economy, Corporate culture and

economic theory: Cambridge University Press.

Lee, Kim J. S. and Kelvin Yu. 2004. “Corporate culture and organizational performance.”

Journal of Managerial Psychology, 19(4): 340–59.

Page 45: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

44

Miller, Danny. 1991. "Stale in the saddle: CEO tenure and the match between organization and

environment." Management science 37(1): 34-52.

Moniz, Andy. 2015. “Inferring employees’ social media perceptions of goal-setting corporate

cultures and the link to firm value.” Unpublished Working Paper.

Nieva, V. F. and J. Sorra. 2003. “Safety culture assessment. A tool for improving patient safety

in healthcare organizations.” BMJ Quality & Safety, 12(suppl 2): ii17-ii23.

http://qualitysafety.bmj.com/content/qhc/12/suppl_2/ii17.full.pdf.

O’Reilly III, Charles A., David F. Caldwell, Jennifer A. Chatman, and Bernadette Doerr.

2014. "The promise and problems of organizational culture: CEO personality, culture, and

firm performance." Group & Organization Management, 39(6): 595-625.

O'Reilly, Charles A., Jennifer Chatman, and David F. Caldwell. 1991. “People and

organizational culture. A profile comparison approach to assessing person-organization fit.”

Academy of management journal, 34(3): 487–516.

O'Reilly, Charles A. and Jennifer A. Chatman. 1996. Research in Organizational Behavior.

18: 157-200, Culture as social control: Corporations, cults, and commitment.

Ott, J. Steven. 1989. The organizational culture perspective. Dorsey Press

Popadak, Jillian A. 2013. “A corporate culture channel. How increased shareholder governance

reduces firm value.”

Quinn, Kevin M., Burt L. Monroe, Michael Colaresi, Michael H. Crespin, and Dragomir R.

Radev. 2010. "How to analyze political attention with minimal assumptions and costs."

American Journal of Political Science 54(1): 209-228.

Rivera, Lauren A. 2012. "Hiring as cultural matching: The case of elite professional service

firms." American Sociological Review 77(6): 999-1022.

Schein, Edgar H. 1992. Organizational Culture and Leadership. San Francisco: Jossey-Bass

Publishers.

Schwartz, Shalom H. 1992. “Universals in the content and structure of values. Theoretical

advances and empirical tests in 20 countries.” Advances in experimental social psychology,

25: 1–65.

Serfling, Matthew A. 2014. "CEO age and the riskiness of corporate policies." Journal of

Corporate Finance 25 (2014): 251-273.

Shapiro, Samuel S. and Martin B. Wilk. 1965. “An analysis of variance test for normality

(complete samples).” Biometrika, 52(3/4): 591–611.

Sheridan, John E. 1992. “Organizational culture and employee retention.” Academy of

management journal, 35(5): 1036–56.

Sørensen, Jesper B. 2002. “The Strength of Corporate Culture and the Reliability of Firm

Performance.” Administrative Science Quarterly, 47(1): 70.

Yoon, Byungun, and Yongtae Park. 2004. "A text-mining-based patent network: Analytical

tool for high-technology trend." The Journal of High Technology Management Research 15(

1): 37-50.

Page 46: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

45

Appendix

Figure A1: Example of Employee Reviews from Glassdoor.com

Page 47: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

46

Figure AF2: Word Clouds. Generated by LDA with 15 topics and

30 words each

Page 48: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

47

Figure AF3: Kernel Density Innovation

Figure AF4: Kernel Density Stability

Page 49: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

48

Figure AF5: Kernel Density Outcome Orientation

Figure AF6: Kernel Density Teamwork

Page 50: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

49

Figure AF7: Kernel Density Detail Orientation

Figure AF8: Kernel Density Respect for People

Page 51: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

50

Figure AF9: Kernel Density Aggressiveness

Figure AF 9: Distribution of Differences to Sector Average for OCP Dimensions: Innovation,

Stability, Outcome Orientation, Teamwork

Page 52: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

51

Figure AF 10: Distribution of Differences to Sector Average for OCP Dimensions: Respect for People,

Detail Orientation, Aggressiveness, and the combined difference for all seven dimensions

Page 53: Corporate Performance and Cultureconference.iza.org/conference_files/DATA_2018/pasch_s26796.pdf · economics into the increasingly important world of Big Data”. They apply computational

52

TABLE AT1

Dimensions and Attributes of the Organizational Cultural Profile

Dimension Associated Attributes

Innovation Innovation, Opportunities, Experimenting,

Risk Taking, Careless

Stability Rule oriented, Stability, Predictability,

Security

Respect for People Fairness, Tolerance, Respect for

Individual

Outcome Orientation Achievement oriented, Action oriented,

High Expectations, Results Oriented

Attention to Detail Precise, Attention to Detail, Analytical

Team Orientation Team Orientation, Collaboration, People

Oriented

Aggressiveness Competitive, Aggressive, Socially

irresponsible The Table shows the seven dimensions of the OCP. O'Reilly, Chatman, and Caldwell (1991) determine

for each dimension correlated attributes and values, which are shown in column “Associated Attributes”