People Management Skills, Employee Attrition, and

Manager Rewards: An Empirical Analysis∗

Mitchell Hoffman Steven Tadelis

U. Toronto Rotman & NBER UC Berkeley Haas & NBER

February 2018

Abstract

How much do a manager’s interpersonal skills with subordinates, what we call people
management skills, affect employee outcomes? Are managers rewarded for having such
skills? Using personnel data and surveys of employees about their managers at a large,
high-tech firm, we show that survey-measured people management skills have a strong
negative relation to employee turnover. A causal interpretation is reinforced by research
designs exploiting new workers joining the firm and manager moves. However, people
management skills do not consistently improve non-attrition outcomes. Better people
managers are themselves more likely to receive higher subjective performance ratings
and to be promoted.

JEL Classifications: M53, D23, J24, L23

Keywords: People management, attrition, supervisors, employee surveys, productivity

∗
We thank Camilo Acosta-Mejia, Jordi Blanes i Vidal, Nick Bloom, Kevin Bryan, Laura Derksen, Wouter
Dessein, Guido Friebel, Maria Guadalupe, Tom Hubbard, Pat Kline, Harry Krashinsky, Eddie Lazear, Bentley
MacLeod, Orie Shelef, Chris Stanton, and especially Kathryn Shaw, as well as numerous conference/seminar
participants for helpful comments. We are grateful to the anonymous firm for providing access to proprietary
data. We also thank several managers from the firm for their insightful comments. One of the authors has
performed paid work for the firm on topics unrelated to HR and the workforce. The paper was reviewed to
ensure that confidential or proprietary information is not revealed. Hoffman thanks the Stanford Institute for
Economic Policy Research for its hospitality. Hoffman acknowledges financial support from the Connaught
New Researcher Award and the Social Science and Humanities Research Council of Canada.
1 Introduction

The relationship between managers and employees is fundamental to the success of firms.
While a growing body of research in labor and organizational economics examines the role
of management practices in understanding large productivity differences across firms and
countries (Ichniowski et al., 1997; Bloom and Reenen, 2007; Bloom and Van Reenen, 2011;
Bloom et al., 2017), less attention has been devoted to managers themselves. It seems evident
that good managers matter. Many people pay handsomely to attend business school to become
better managers, and scores of books are written every year on how to become a better
manager. Unfortunately, little empirical evidence exists regarding the managerial production
function and the influence of managers on their employees, particularly regarding the influence
of interpersonal skills (or what is sometimes called “people management”).
How much does good people management by managers matter? Is good people man-
agement rewarded inside the firm? We seek to answer these and related questions using rich
employee surveys conducted by a multinational technology and services firm. Employees in
our firm are asked to evaluate their managers on a number of dimensions, e.g., whether they
are trustworthy or whether they provide adequate coaching. We use these surveys to measure
each manager’s people management skills.
Progress has been made recently in examining how much managers matter using a “value-
added” (VA) approach. Bertrand and Schoar (2003) examine how much CEOs matter for
various decisions in firms by regressing various firm outcomes on CEO fixed effects, pioneering
an approach that has been pursued by a large subsequent literature in finance. In a recent
important paper, Lazear et al. (2015) use data from one firm to examine to what extent low-
level managers (specifically, frontline supervisors) matter for productivity, finding that they
matter a great deal. Bender et al. (2018) analyze interactions between employees/managers
and management practices in Germany. Hoffman et al. (2018) use data from several low-skill
firms to examine the determinants of frontline manager productivity around the world.

1
 While these studies are of great interest, the VA approach faces two main limitations.
First, VA studies require good objective data on worker productivity. However, in many
firms, direct data on individual worker productivity is often scarce, and sometimes impossible
to measure, particularly in high-skill, collaborative environments. When data are available,
productivity metrics may be subject to various shocks (e.g., business generated by a law firm
partner could be adversely affected by the exit of a single prominent client, who decided
to leave the firm for reasons having nothing to do with the partner). Second, VA estimates
provide researchers with the overall impact of a given manager on individual outcomes, not the
separate impact of people management skills. A manager may appear to have desirable fixed
effects for various reasons separate from subordinate-related interpersonal skills, such as ability
to bring in high-value clients, thereby making his or her employees look more productive.
We take a different (and complementary) approach. Rather than calculate VA, we create
a measure of people management skills using employees’ survey responses about their manager.
We then explore the extent to which people management skills relate to employee outcomes,
with the greatest focus on employee attrition. The data from the firm cover thousands of
managers and tens of thousands of employees. In high-tech firms, employee turnover is believed
to be a key way by which knowledge and ideas are acquired and lost (e.g., Shankar and
Ghosh, 2013; Stoyanov and Zubanov, 2012). As such, many high-tech firms, including ours,
are deeply interested in what can be done to reduce turnover, particularly turnover of their
highest performers. Our approach is complementary to VA because it answers how much
subordinate-related interpersonal skills matter (instead of how much managers matter overall).
A central task for managers is to enhance the productivity of their employees and to
help them succeed in their jobs. Asking employees about what managers do to improve
their performance thus seems like a natural way to measure people management skills, and
is one that has been pursued by many firms. Brutus et al. (2006) report that over 1/3 of
US and Canadian organizations in their survey use “multi-source assessments” (as opposed
to assessing individuals based on their managers) and Pfau et al. (2002) report that 65% of

2
firms use 360-degree performance evaluations.1 Indeed, our approach of analyzing managers
using employee surveys thus appears to align closely to the data practices of many firms.
There are several challenges in using employee surveys to measure people management.
First, there is concern about non-response bias, but the response rate at our firm is over 95%.
Second, one may worry that employees may not be truthful, e.g., workers may not want their
manager to know that they evaluated them negatively. This concern is mitigated due to the
confidential nature of the survey. Workers are told truthfully that their individual responses
will never be observed by the firm. Instead, managers receive aggregated results, and only
for managers with a minimum number of employees responding. Our data is thus limited to
manager-year averages for various qualities ascribed to them by their employees. This feature
protects worker confidentiality, but does not limit our analysis, given our focus on under-
standing behavior at the manager level. Third, survey responses may contain measurement
error for several reasons, e.g., inattentiveness, sampling error, or different employees treating
different questions differently. We address this using an instrumental variable (IV) strategy
where a manager’s score in one wave is instrumented using his or her score in the other wave.
Our main finding is that people management has a strong negative relation to employee
attrition. Our main IV results imply that increasing a manager’s people management skills
from the 10th to 90th percentile predicts a 60% reduction in turnover. These results are quite
strong in terms of retaining the firm’s high performers, both defined in terms of classifying
employees based on persistent subjective performance score differences and using the firm’s
definition of regretted voluntary turnover. Beyond classical measurement error, IV addresses
measurement error in people management that is contemporaneously correlated with attrition.
Still, the question remains whether these results are causal. Even for our IV estimates,
there are concerns about non-contemporaneous measurement error in measured people man-
agement skills that is correlated with employee attrition, as well as concern that the firm is
optimally sorting managers or employees together. We address this concern using multiple
1
Prominent examples of firms using such surveys include Google (Garvin et al., 2013) and Royal Bank of
Canada (Shaw and Schifrin, 2015), as documented in business case studies.

3
identification strategies, some of which are in the spirit of those in the teacher VA literature
(Chetty et al., 2014). Our first strategy analyzes outcomes of employees who join mid-way
through our sample, using a manager’s quality measured before an employee joins the firm as
an instrumental variable. This addresses concern about non-permanent, unobservable shocks
affecting turnover and manager ratings, and reduces concern that the results are driven by
the firm sorting managers and employees based on long-time information about the employee.
Our second strategy additionally analyzes instances of workers switching managers, allowing
us to test for non-random assignment of managers and workers, and to analyze how the impact
of people management skills varies based on time together between a manager and employee.
Our third strategy exploits managers moving across locations or job functions within the firm,
allowing us to address more permanent unobservables (beyond what are already measured us-
ing our rich, baseline controls), as well as to rule out assignment bias. All strategies support
people management skills having a strong, causal effect on attrition.
Having examined attrition, we also examine the relation between people management
and other employee outcomes. Interestingly, we do not find a consistent relation between
people management skills and employee subjective performance, salary growth, or probability
of promotion, at least once employee fixed effects are controlled for. Thus, good people
management affects some outcomes (namely different attrition variables), but not others.2
Our secondary result is that managers with better people management skills get “re-
warded” by the firm. Better people managers attain substantially higher subjective perfor-
mance scores and are more likely to be promoted. Such results are consistent with the firm
valuing the role of good people management skills in reducing employee turnover.
Our paper contributes to several literatures. First, it is related to other work on the
importance of individual managers. In addition to work on VA (mentioned above), Bandiera
et al. (2017) classify CEOs into two types using machine learning techniques and time use
data, finding that one type (representing a higher tendency to delegate) outperforms the
2
We also observe a positive relation between people management and employee engagement, but it is also
not consistently statistically significant.

4
other. Friebel et al. (2017) conduct a field experiment where they send a letter from the
CEO to store managers explaining that reducing turnover is important; they find that this
intervention reduced turnover by 1/3.3 Most related to our paper is the above-cited work
of Lazear et al. (2015), who study the impact of lower-level managers in a low-skill setting.
Relative to existing work, our paper contributes by asking a different question (i.e., what is
the impact of people management skills as opposed to the impact of managers overall) and
using a different methodology (i.e., to measure manager skills using employee surveys).
Second, it relates to work in general on knowledge-based employees. Much of empirical
personnel economics focuses on relatively low-skilled jobs (e.g., truckers, retail, and farm-
workers), partially because it is often relatively simple to measure individual productivity.4
In contrast, for high-skilled jobs for knowledge employees, production is often complex, multi-
faceted, and involves teamwork. Our analysis sheds light on the managerial production func-
tion in such a high-skilled setting.
Third, it is related to work on subjective performance evaluation and workplace feed-
back. Employee surveys bring to bear an advantage often ascribed to subjective performance
evaluation, namely, that they help account for difficult-to-measure aspects of performance
(Baker et al., 1994b). Our paper brings forward a new aspect of performance evaluation,
namely reports from a manager’s employees, that has not been previously explored in eco-
nomics.5 Fourth, it relates to studies of compensation and reward within organizations (e.g.,
3
In addition, Glover et al. (2017) show that employees perform worse when paired with more ethnically
biased managers. Frederiksen et al. (2017) show that workers benefit financially from having a supervisor who
is lenient in performance evaluations.
4
For notable exceptions in personnel economics also using high-skilled workers, see, e.g., Bartel et al. (2017);
Kuhnen and Oyer (2016); Burks et al. (2015). Beyond studying high-skilled workers, our choice of studying
managerial skills within a firm in the high-tech sector seems particularly relevant, as such firms are likely to
be significant contributors to growth in the 21st century (Aghion et al., 2017).
5
There is economics work using satisfaction surveys (which sometimes include questions about supervisors)
showing that more satisfied workers are less likely to attrite (e.g., Clark, 2001; Frederiksen, 2017). In contrast,
we focus specifically on managers, while using research designs aimed at measuring the impact of underlying
people management skills (as opposed to a person’s individual perception of their job). There is also economics
work on student evaluations of teachers (e.g., Beleche et al., 2012). Carrell and West (2010) show that
teacher evaluations positively correlate with contemporaneous student VA, but negatively correlate with later
achievement. There is also a vast psychology literature on leadership (Yukl, 2010) that includes 360 degree
evaluations (e.g., Atkins and Wood, 2002). However, this work seems primarily focused on different issues (e.g.,
consistency of ratings) as opposed to the causal impact of people management skills on employee outcomes.

5
Baker et al., 1994a), providing novel evidence that people management skills are rewarded
within the firm.
Section 2 describes the data. Section 3 describes our empirical strategy. Section 4 pro-
vides our main analyses on how managers’ people management skills affect employee attrition.
Section 5 analyzes how people management skills affect outcomes other than attrition. Section
6 analyzes to what extent people management is rewarded. Section 7 concludes.

2 Data and Institutional Setting

Our data, obtained from a technology and services company, cover a period of two years
and five months, some time during 2011-2015. To preserve firm confidentiality, certain details
regarding the firm and exact time period cannot be provided. We refer to the three years of the
data as Y1 , Y2 , and Y3 . The data are organized as a worker-month panel. Between January
Y1 and May Y3 , we observe several dozen thousands of employees and several hundreds of
thousands of worker-months.
The firm is divided into several broad business units (similar to industry). In addition,
workers are classified by job function (similar to occupation). A core job function at the firm
is engineering (comprising 36% of worker-months in our sample), and there are also many
workers in various non-engineering functions (e.g., marketing, finance, sales). Furthermore, as
in many high-tech companies, the firm has a large number of lower-skilled workers in customer
service / operations, but we exclude them from our analysis, given our broad motivation of
better understanding high-skilled workplaces.6
About 21% of observations (worker-months) are filled by individuals in managerial roles,
so the majority of observations are for non-managers (often referred to in industry as individual
contributors). While our data begin in Jan. Y1 , the majority of the workers are hired before
6
In the dataset we were provided, over 1/3 of observations are from customer service workers, making them
even more numerous than engineers. Thus, if we did not exclude customer service workers, they would play
an outsized role in our analysis. Customer service workers also play a qualitatively different role from most
other workers at the firm. Still, our main results in Table 3 are robust to including customer service workers.

6
that date. Still, 38% of the employees in the data were hired on or after Jan. Y1 . The
data cover workers only, not applicants. We next provide information on employee outcomes,
manager assignment, and the employee surveys, with further details regarding the data in
Appendix B.

2.1 Employee Outcomes

In knowledge-based firms such as the one we study (as well as in many non-knowledge-based
firms), employee performance often has multiple dimensions, and is rarely defined according to
a single productivity metric (e.g., output per hour). There are several core employee outcomes
in our data, the most important one for our purposes being employee turnover:

• Turnover. Turnover is a significant issue in many firms, particularly in high-tech. High-

tech firms are often keenly interested in reducing turnover because employee knowledge
is a key asset and turnover is an important means of knowledge transfer.7 This is
particularly true for employees in states like California, where non-compete agreements
are unenforceable and firms fear losing ideas to competitors (Balasubramanian et al.,
2017). We separately observe dates of voluntary attrition (“quits”) and involuntary
attrition (“fires”). We also observe whether the firm classified quits as regretted (i.e.,
wished it had avoided) or non-regretted, with further information on this classification
in Appendix B. Thus, the data allow us to go much further than is typical in economics
in analyzing turnover.

• Subjective performance. The firm’s subjective performance scores are on a discrete

1-5 scale, as is the case in many firms using subjective performance evaluation (e.g.,
Frederiksen et al., 2017), and are set biannually. The scores are set in a process involving
an employee’s immediate manager as well as higher-up managers. While there are some
broad guidelines for the distribution of subjective performance scores across various
7
In fact, as is common in many firms, our firm has analysts who try to predict and reduce turnover.

7
 units within the firm, there is not a fixed “curve” across managers in the number of
subjective performance scores that can be provided.8

• Salary increases. While it is difficult to measure the productivity of knowledge work-

ers, one way of proxying productivity improvements is the extent to which an employee’s
salary increases (or to proxy productivity using the level of salary). Because salaries are
listed in local currency, we restrict analysis of salary to employees paid in US dollars
following Baker et al. (1994a).

• Promotions. Promotions are pre-defined in the data provided by the firm, and cor-
respond roughly to an increase in a person’s salary grade. Another recent paper using
promotions as a proxy for knowledge worker productivity is Brown et al. (2016).

While these outcomes are all important and capture valuable aspects of worker performance,
we do not claim that we are fully measuring “productivity” in our setting.9 Certain aspects of
productivity, such as the contributions of a businessperson to a new marketing strategy or the
value of an engineer’s computer code, seem inherently difficult to quantify. Indeed, beyond
being hard to measure, productivity in some knowledge jobs seems even hard to define. This
increases the usefulness of using surveys to measure manager quality (relative to using VA).
Different employee outcomes are available at different frequencies, but are coded in our
data at the monthly level. Attrition and promotion events are coded in our data at the
monthly level using exact dates for these events. Subjective performance reviews occur twice
per year, but are also coded month-by-month. Annual salary is tracked at the monthly level,
though salary increases are more likely to occur during two particular months.
8
At high levels of aggregation within the firm (i.e., for top managers), there may be a curve with respect
to subjective performance. To address possible bias from curving, we verified that our subjective performance
results are robust to excluding high-level managers (with “high-level” as defined in Appendix Table C18).
9
In addition to these outcomes, we observe employee engagement. Engagement is a number from 0-100
about how engaged the employee is feeling (via the same survey that is used to elicit employees’ perceptions
of their manager), and is defined using various job satisfaction questions. We do not emphasize engagement,
one, because it is not measured at the individual level (and is only available at the manager level), and two,
because it is measured using the same survey as people management skills.

8
2.2 Assignment of Managers to Employees

Managers usually manage employees within their function and line of business, and this is
reflected in the initial assignment of employees to managers. Assignment of employees to
managers reflects the projects and functions that require employees at any given time. Geo-
graphic needs also dictate the circumstances in which employees may experience a change of
manager. The company has an online system where managers post internal workforce needs,
and new employee-manager matches can form based on these online postings. Managers play
a key role in hiring and are also involved with dismissals. Thus, it is clear that employees
at the firm are not being randomly assigned to different managers. Instead, managers play a
significant role in selecting employees for their teams. We further discuss manager assignment
at the start of Section 3, as well as in Section 4.
On average, across managers in our sample, a manager manages about 9 employees at
one time. However, the average number of employees per manager is 11 when managerial
span is weighted by employee-months. In our sample, employees experience an average of 1.4
managers during the dataset (and 1.5 managers when managers per employee is weighted by
employee tenure). These numbers reflect that, in our sample, we do not count worker-months
if the manager’s team at survey time is too small to have the survey administered.10

2.3 Employee Surveys

Every year, employees are given a detailed survey. The goal of these types of surveys is for the
firm’s Human Resource (HR) department and executives to gain an accurate sense of employee
opinions. Because the surveys are designed to ensure the anonymity of responses, survey
information about one’s managers is only collected on managers who manage a minimum
number of individuals.11 Thus, workers on small teams (i.e., where team size is below the
10
If we did not make the restriction that a manager’s people management score be non-missing in the current
and other period, there would be an average of 2.3 managers per worker (see Appendix Table C1).
11
In the first year of the survey (Y1 ), the threshold was 3 employees, whereas in the second year of the
survey (Y2 ), the threshold was 5 employees. Technically, the survey is “third-party confidential” instead of
“anonymous,” according to the firm. Anonymous means that it would be totally impossible to tie responses

9
survey threshold) are not part of the analysis, but our results are robust to imputing survey
scores for the managers of small teams. Appendix B gives further details.
Surveys of this type are typically administered before year-end, and consistent with this
industry norm, the surveys in our data were performed in September in Y1 and Y2 . The survey
had the same format and same manager questions in Y1 and Y2 .
For our main analysis, to match outcomes with their associated survey, observations from
January Y1 -September Y1 are assigned the survey information from the Y1 survey, whereas
other observations are assigned the survey information from the Y2 survey.12 For both the Y1
and Y2 surveys, the response rate was 95%.
We also have data for a third survey administered in Sept. Y3 , but we don’t use it for
our main analysis for three reasons. First, the survey format changed. Second, many prior
questions were removed or replaced, including some of the manager questions. Third, the
survey was not administered to one large business unit, so there are many missing values.

Manager questions. Various survey questions are asked every year about what em-
ployees think about their managers. For each question, employees are asked whether they
Strongly Disagree, Disagree, Neither Agree nor Disagree, Agree, or Strongly Agree. Specifi-
cally, we observe answers to the following survey items:13

1. My immediate manager communicates a clear understanding of the expectations from

me for my job.

2. My immediate manager provides continuous coaching and guidance on how I can improve
my performance.

3. My immediate manager actively supports my professional/career development.

4. My immediate manager consults with people for decision making when appropriate.
to employee attributes. Third-party confidential means the survey vendor, a third-party independent firm,
has access to responses so it can tie them to employee attributes to generate statistical information.
12
This is how firm analysts assigned the scores in the cleaned data provided.
13
To preserve firm confidentiality, the wording may be slightly modified from the original.

10
 5. My immediate manager generates a positive attitude in the team, even when conditions
are difficult.

6. My immediate manager is someone whom I can trust.

A manager’s rating on an item is measured as the share of employees who marked Agree or
Strongly Agree.14 For example, if a manager has 8 direct reports, and 6 of them marked Agree
or Strongly Agree for one of the items, the manager’s score on that item would be 75 out of
100 in the data provided to us. If employees experience multiple managers over the survey
period, they only rate their most recent manager.
A manager’s overall rating (MOR) is the average of scores on the 6 items. For example,
if a manager had score of 100 on the first 3 items and a score of 50 on the second 3 items, the
manager’s MOR is 75. The MOR is easy to compute and is used by the firm in its internal
reporting and communications. We will use MOR as our main measure of employee-survey-
based manager quality, and discuss this further below in Section 2.5.
The manager questions are toward the end of a longer survey covering many topics (e.g.,
general worker engagement and satisfaction).

2.4 Sample Creation and Summary Statistics

To create our sample, we restrict attention to worker-months where an employee has a manager
with a non-missing MOR for the current period, as well as a non-missing MOR in the other
period (we define “periods” below in Section 3.1). This sample restriction is required for our
IV analysis, where we instrument manager MOR in the current period using MOR in the
other period. We also exclude April and May of Y3 from our sample, as the firm’s location
14
This is how the data are prepared by the third-party survey provider (presumably in part to protect
anonymity of responses), and is thus also the form that the firm uses in its internal reporting. Therefore, it is
impossible for us to analyze other moments of the survey responses (e.g., the standard deviation of responses
about a manager). However, to our understanding, it is common practice in such surveys to break up the
5-answer scale into 2 or 3 parts. For example, exhibit 7 of Garvin et al. (2013) suggests that Google grouped
the 5 answers into Unfavorable (Strongly Disagree or Disagree), Neutral (Neither Agree nor Disagree), and
Favorable (Agree or Strongly Agree) in its own people management survey. Thus, using the share of employees
marking 4 or 5 (as we do) seems consistent with how many firms measure their managers on similar surveys.

11
identifiers change in these months compared to before. Thus, our sample runs from January
Y1 -March Y3 (though our main results are qualitatively similar to extending through May Y3 ).
Table 1 provides summary statistics for our sample. The employee attrition rate is 1.37%
per month (or about 15% per year). The majority of separations are voluntary (“quits”), but
there are still a sizable number of involuntary separations (“fires”). There are a number of
exits which are not classified in the data as voluntary or involuntary.
The average MOR is about 81 out of 100. About 81% of employees are co-located
with their manager, whereas the remainder are managed remotely. While the number of
observations cannot be shown to preserve firm confidentiality, note that observation counts
vary slightly by variable, reflecting challenges in linking together many firm datasets.
Appendix Table C1 provides summary statistics, but before imposing a manager’s MOR
be non-missing in the current and other periods. Appendix A.1 discusses further.

2.5 Properties of MOR

Is MOR the right measure? Is there another way of combining the six manager questions
that is more sensible than a simple average? We explore this using principal component anal-
ysis. As seen in Appendix Table C3, the first component explains about 70% of the variation
in manager scores. Interestingly, the first component is quite close to an equally weighted
average of the 6 individual items. Thus, beyond being very simple, another justification for
using MOR is that it is close to the first principal component of the six questions, a component
that explains a large share of the variance.

Persistence. Figure 1 shows that manager scores are somewhat persistent over time
using a binned scatter plot with no controls. The coefficient of 0.37 means that a manager
who scores 10 points higher in the Y1 survey in MOR is scored 3.7 points higher on average
in the Y2 survey. Appendix Figure C1 shows that similar correlations are seen on each of the
six manager questions. Table 2 observes a similar pattern of some persistence while applying
control variables (and using normalized survey scores). Each column takes MOR or one of the

12
managerial quality questions from the Y2 survey. The score is then regressed on the various
manager quality questions from the Y1 survey. For example, column 1 shows that a manager
who performs 1σ higher in Y1 scores 0.37σ higher in Y2 , conditional on controls.
The predictiveness of the scores over time is sizable, but perhaps not as high as some
readers might expect. We believe that the main reason is measurement error in the surveys.
Even though the firm strongly encouraged workers to take the surveys quite seriously (which
is reflected in the 95% response rate), measurement error often occurs when respondents are
asked to answer many questions, particularly subjective ones (Bound et al., 2001). Mea-
surement error could arise from inattention or survey fatigue, as well as factors that could
influence how one rates their manager separate from true underlying manager quality (e.g.,
worker mood). Further, measurement error can arise from the fact that MOR scores are cre-
ated by taking the share of individuals marking Agree or Strongly Agree to a question, thereby
introducing noise from an average over a discrete categorization.15 Section 3 details how our
empirical strategy addresses various challenges from measurement error in the survey.
Despite the likely presence of measurement error, the Table 2 results are consistent with
the view that managers have particular characteristics that are somewhat persistent over time.
One challenge with this interpretation is that the various manager characteristics are correlated
with one another.16 To address this, we also regress each current characteristic on all the six
past characteristics at once. Appendix Table C5 shows that each individual characteristic
predicts the future characteristic even while controlling for the other characteristics.17
As noted above in Section 2.3, we don’t use the Y3 survey for our main analysis because
the format changed and it was not administered to a large business unit. Still, Appendix
15
Other factors beyond measurement error may also limit persistence. First, a manager’s responsibilities,
tasks, and projects change over time. A manager might be perceived as providing excellent coaching and
guidance for one type of project, but not for another. Second, if a manager scores badly multiple years in a
row, he/she is invited to attend a “bootcamp” to improve manager effectiveness. Appendix Table C4 shows
a transition matrix for MOR quintiles.
16
The correlation is relatively high, though still much less than 1. See Appendix Table C2.
17
Another concern with interpreting managerial characteristics as relatively persistent is that manager scores
could reflect persistence of worker characteristics or how workers answer the survey as opposed to manager
characteristics. Thus, we have also repeated Table 2 while restricting attention to managers who move locations
across the firm in the second period. In this robustness check, we continue to see substantial persistence of
managerial characteristics across surveys.

13
Table C6 shows that the result on the persistence of overall MOR (column 1 of Table 2) is
qualitatively robust (but smaller) when including the Y3 survey.

3 Empirical Strategy

While there are several parallels between the teacher VA literature and our analysis, there
are also key differences. First, we are estimating the impact of a survey-measured regressor
(namely, survey-measured people management skills) instead of estimating manager fixed
effects. Thus, we need to address the important issue of measurement error in our survey (as
opposed to sampling error in estimating large numbers of fixed effects).
Second, unlike in schools (where teachers generally do not choose their students), man-
agers at our high-tech firm play a critical role in hiring and selecting people for their team.
Indeed, practitioners frequently argue that one of the most important parts of being a good
people manager is selecting the right people (Harvard Management Update, 2008). Even if we
could convince our firm to randomly assign employees to managers, such an experiment would
not be informative of the overall impact of good people management skills since it would rule
out better people managers selecting better people. Rather, differences in employee qual-
ity across managers might be viewed as a mechanism by which managers improve employee
outcomes as opposed to a source of bias.
A more informative hypothetical experiment for our setting (and one we try to approx-
imate in our design based on managers switching locations or job functions in Section 4.4),
would be to randomly assign managers to different parts of the firm and then observe employee
outcomes, thereby reflecting the role of managers in selecting and motivating their teams.18
Still, even if we do not wish to rule out better managers selecting better people, we need to
address the possibility of the firm optimally sorting managers and employees together, which
we discuss further below.19
18
One would also ideally randomly assign differing levels of people management skills to managers.
19
Beyond a manager selecting employees for his or her team, and beyond the firm sorting employees and
managers, another situation that could arise is one where a person has a reputation as a great people manager.

14
3.1 Econometric Set-up

We wish to estimate how much the underlying people management skill of manager j, mj ,
affects an outcome, yit , of employee i:

yit = βmj(i,t) + εit (1)

where j(i, t) represents that j is the manager of employee i at time t, though we will hence-
forth abbreviate j(i, t) simply by j. Given we are using an imperfect survey, one concern is
that people management skill, m, is measured with error; instead of true underlying people
management, we only observe the noisy survey measure, m.
e In our data, we have the two
main waves of the survey, giving us two manager scores m
e 1 and m
e 2 , with m
e j,τ = mj + uj,τ ,
τ ∈ {1, 2}. In our data, t is at the monthly level, whereas there are two values of τ .
Perhaps the simplest approach to analyzing the impact of people management skills is
to estimate OLS regressions of the form:

yi,t = bm
e j,τ (t) + θi,t (2)

where θi,t is an error term; and where τ (t) = 1 if t ≤ month 9 of Y1 and τ (t) = 2 if t >
month 9 of Y1 .20 We refer to τ as the period. However, OLS models may be biased by
measurement error. An alternative approach (e.g., Ashenfelter and Krueger, 1994; Ashenfelter
and Rouse, 1998) is to instrument one survey measure with the other one:

yi,t = bm
e j,τ + θi,t (3)

m
e j,τ = cm
e j,−τ + ηj,t

where m
e j,−τ is the measured people management score of manager j in the period other than
the current one, and θi,t and ηj,t are error terms.
Good candidates may try to sort onto a good manager’s team. While we try to reduce concerns about sorting
below, it is not clear that one would want to control for this. Indeed, having a “brand” and thus attracting
strong hires is a key way good people management may matter, particularly in high-skilled settings, such as
the one we study.
20
That is, τ (t) corresponds calendar months to survey periods. Recall that the Y1 survey was administered
in m9 of Y1 . Thus, we analyze employee outcomes during period 1 as a function of their manager’s rating
during period 1. This assignment of calendar dates to survey periods is also used by the firm for their internal
reporting.

15
 Instead of assuming that the measurement error is classical, we will consider the possibil-
ity that the measurement error could be correlated with unobserved determinants of employee
outcomes, e.g., that being on a good project could affect how an employee rates their manager,
as well as whether that employee attrites. That is, compared to most empirical studies where
there is measurement error, we make fewer assumptions. However, like most studies, we do
assume that measurement error is uncorrelated with a manager’s true people management
skill:21

Assumption 1 cov (mj , uj,τ ) = 0 for τ ∈ {1, 2}.

While we do not expect Assumption 1 to be literally true (given that there are caps of the
management score at 0 and 100), we believe that it is approximately true in our setting,
particularly because people are not selecting their own management score.22
We now compare OLS and IV estimators for this setting. For ease of exposition, we
cov(yt ,m
eτ)
suppress i and j subscripts. For OLS, we use plim(bbOLS ) = var(meτ)
plus Assumption 1 to
get the below (derivation in Appendix A.2):
σ2 cov(εt , uτ ) cov(εt , m)
plim(bbOLS − β) = − 2 u 2 β + + (4)
σ +σ σ2 + σ2 σ2 + σ2
| m{z u } | m {z u } | m {z u }
Attenuation Bias Contemp. Corr. ME Assignment Bias

In (4), the first term or Attenuation Bias, is standard under OLS when there is classical
measurement error on the right-hand side. In the second term or inconsistency from Con-
temporaneously Correlated Measurement Error, the numerator, cov (εt , uτ ), is the covariance
between the measurement error from the survey and unobservables that affect employee out-
comes. We believe that such measurement error is likely to be positive, but that is not
necessarily the case. For example, one issue for analyzing attrition as an outcome is that
there are individuals who quit before they get to take the survey. A manager may appear to
have a better score on the survey than if the employee was allowed to take part in the survey.
21
Our analysis of measurement error draws heavily (in content and notation) from Pischke (2007).
22
Assumption 1 seems most likely to be systematically violated when people are answering surveys about
themselves and there are social pressures such as conformity bias, e.g., someone with a low amount of actual
schooling or earnings might feel social pressure to report that they have more schooling or earnings than they
actually have (as in Bound and Krueger (1991)).

16
In the third term or Assignment Bias, the numerator, cov(εt , m), represents the correlation
of worker-level unobservables with manager quality. This could be positive or negative.
Next, consider the IV estimator where we instrument a manager’s score during one period
with the manager’s score in the other period (as in equation (3) above). Note that different
employees may evaluate the same manager during two different periods. Using plim(bbIV ) =
cov(yt ,me −τ )
cov(me τ ,m
e −τ )
, we get:
cov (uτ , u−τ ) cov (εt , u−τ ) cov (εt , m)
plim(bbIV − β) = − 2 β+ 2 + 2 (5)
σm + cov (uτ , u−τ ) σm + cov (uτ , u−τ ) σm + cov (uτ , u−τ )
| {z } | {z } | {z }
Attenuation Bias Asynchronously Corr. ME Assignment Bias

As for OLS, the expression for the consistency of IV (equation 5) has three terms. The first
term has cov (uτ , u−τ ) in place of σu2 . Thus, if the measurement errors are uncorrelated across
the two surveys, there is no attenuation bias. This assumption seems reasonable for certain
types of measurement error, such as sampling error due to small numbers of subjects, one-time
data imputation, or people being in a good mood because the current project is going well.
Other types of measurement error might be more persistent, e.g., there might be persistence
if people on a manager’s team have a general tendency to rate managers highly on surveys.
As we discuss later, such correlations can be avoided by looking at managers who are rated
by mostly different employees in the first and second periods, e.g., managers who move across
locations in the firm. In such circumstances, we would expect substantially less attenuation
bias than in OLS.
The second term of (5) has cov (εt , u−τ ) instead of cov (εt , uτ ) in the numerator. That
is, it involves the covariance between measurement error in the other period and the unob-
served determinants of performance in the current period (instead of the covariance between
measurement error in the current period and the performance equation error in the current
period). For measurement error due to inattention or non-response, this correlation may be
quite small or zero. For things like being on a good project, this correlation may depend on
how persistent is the shock over time.
The third term of (5) still has cov (εt , m) in the numerator, but it is divided now by

17
 2 2
σm + cov (uτ , u−τ ) instead of σm + σu2 . Thus, IV can amplify assignment bias if cov (uτ , u−τ ) <
σu2 .
We will also present reduced form results, i.e., OLS regressions of yt on m
e −τ :
σ2 cov(εt , u−τ ) cov(εt , m)
plim(bbRF − β) = − 2 u 2 β + +
σ +σ σ2 + σ2 σ2 + σ2
| m{z u } | m {z u } | m {z u }
Attenuation Bias Asynchronously Corr. ME Assignment Bias

Relative to the IV, a disadvantage of the reduced form is that there is still the same attenuation
2
bias as for OLS. A potential advantage is that assignment bias is scaled by σm + σu2 in the
2
denominator instead of σm + cov (uτ , u−τ ).
Throughout the empirical analysis, standard errors are clustered by manager, reflecting
the main level of variation for our key regressor.

3.2 Additional Remarks

• What if a manager’s underlying people management skill varies over time?

This could occur for several reasons, including that a manager’s effort may change over
time; certain managers may be better with some projects or teams than others; and, the
firm may help a manager to improve their people management skills over time. Appendix
A redoes the above formulas allowing manager quality to vary over time. For the IV,
possible attenuation depends now on the size of the covariance of people management
skills over time relative to the covariance of measurement error over time. The OLS and
reduced form expressions are broadly similar to before.

• What happens if people management has persistent effects? The key iden-
tification assumptions for IV are that cov(m
e j,−τ , θit ) = 0 and that the only way that
m
e j,−τ affects yi,t is through its influence on m
e j,τ . One way this can fail is if there are
persistent effects of good people management. Similar to having had a good teacher
in the past, it is possible that good people management could have a persistent effect
over time. We have two responses. First, existing evidence on manager effects suggests
that they are not very persistent: in Lazear et al. (2015), 2/3 of boss effects disappear

18
 after 6 months, and 3/4 disappear after one year.23 Second, some of our identification
strategies rule out persistent effects. For example, when we analyze new workers joining
the firm in period 2, their current manager is interacting with them for the first time at
the firm, so it is impossible that the new workers benefitted from interacting with their
current manager during period 1. Persistent effects can also be addressed by looking at
individuals who switch managers. That we reach qualitatively similar conclusions with
these identification strategies is consistent with people management having primarily a
contemporaneous effect.

• Control variables. While the above setup ignored control variables, adding controls
helps address the possibility that MOR and employee outcomes may differ systematically
in the large firm we study. We control for the firm’s different business units, as well as
for job function (or occupation). We also control for year of hire, current year dummies,
and a 5th order polynomial in employee tenure. We control for location (the firm has
many offices), employee salary grade (or level), and an employee’s manager’s span of
control. Section 4.5 shows that our results are robust to even finer controls.

• Adding fixed effects. Beyond control variables, worker fixed effects can be added to
an IV specification. The key is for some workers to experience multiple managers during
the period for which output is being analyzed. For example, suppose that worker Alan
experiences managers Beth and Collin in period 2. The manager scores of Beth and
Collin will be instrumented with the manager scores that they received in period 1.
23
This evidence differs from ours in several respects. First, we analyze people management skill as opposed
to overall boss effects. Second, we primarily analyze attrition whereas Lazear et al. (2015) primarily analyze
output per hour. Third, we study a high-skill firm, whereas Lazear et al. (2015) study a firm where workers do
a routine job. It is not clear whether such differences would lead to boss effects being more or less persistent,
though we might imagine that the identity and skills of one’s current manager would be particularly important
for our outcome of attrition. Unfortunately, we cannot separately test whether people management is persistent
using our IV strategy.

19
4 Manager Quality and Employee Attrition

Section 4.1 presents our baseline results on the relationship between MOR and employee
attrition. Next, we present our three research designs: new joiners (Section 4.2); new joiners or
employees switching managers (Section 4.3); and managers switching locations or job functions
(Section 4.4). Exploiting different variation (and requiring different identifying assumptions)
and addressing different threats to identification, all three designs provide complementary
evidence supporting that people management skill substantially reduces employee attrition
(Appendix Table C7 summarizes rationales for our different identification strategies). Section
4.5 addresses additional threats to identification and estimates manager VA (in line with past
work on managers such as Lazear et al. (2015)).

4.1 Baseline Results

Panel A of Table 3 shows our baseline results. Column 1 shows a strong first stage (F > 100).
In column 2, the OLS coefficient of -0.154 means that increasing MOR by 1σ is associated
with a monthly reduction in attrition of 0.154 percentage points (hereafter “pp”), which is an
11% reduction relative to the mean of 1.37pp per month. Column 3 presents the IV estimate
where MOR in one period is instrumented with a manager’s MOR in the other period. Here,
the coefficient is substantially larger at -0.475, implying that increasing MOR by 1σ corre-
sponds to a 35% reduction in turnover. By the difference of two Sargan-Hansen statistics,
we reject that the IV and OLS estimates are the same (p < 0.01). That IV is substantially
larger in magnitude than OLS is consistent with OLS being significantly biased downward
in magnitude due to attenuation bias from measurement error.24 Still, as discussed above,
IV could also be biased due to asynchronously correlated measurement error (i.e., measure-
ment error from the non-current period which is correlated with unobserved determinants of
turnover or cov(εt , u−τ ) 6= 0) or assignment bias by the firm (cov(εt , m) 6= 0). We address
24
Our findings suggesting significant measurement error are consistent with Bloom et al. (2017), who doc-
ument significant measurement error in surveys about management practices (instead of about managers).

20
those possibilities further below.
Our IV estimate implies that moving from a manager in the 10th percentile of MOR to
one in the 90th percentile of MOR is associated with a reduction in quitting of roughly 60%
(under the assumption of normality).25 To further assess the IV magnitude, we compare it
to estimates in other studies analyzing turnover, particularly those related to management.
Bloom et al. (2014) show that randomly assigning call-center employees to work from home
reduces turnover by 50%. Thus, having a manager in the 90th percentile of the people man-
agement distribution instead of one in the 10th percentile has an impact on turnover broadly
similar to that of letting employees work from home.26
However, not all attrition is the same. It could be that good managers prevent quits,
but are willing to fire individuals who are not contributing. Thus, Panels B and C perform the
same analyses as Panel A, but separately for quits and fires.27 We observe highly significant
negative IV results for both quits and fires. The coefficient is larger in absolute magnitude for
quits, but is larger in percentage terms for fires, reflecting that fires are rarer in quits. Also,
there is a larger percentage relation between MOR and regretted quits (Panel D) than there
is between MOR and non-regretted quits (Panel E), suggesting that MOR might help reduce
“bad attrition” more than it reduces “good attrition.”
Of course, the distinction between quits and fires is often not clear-cut (e.g., to get
rid of someone, you can ask them to resign), and there is no theoretical distinction in many
models of turnover (e.g., Jovanovic, 1979); further, it is not immediately clear whether it is
“better” for managers to avoid quits or fires. Another way to delve further into turnover is to
25
The 10th percentile of a standard normal is 1.28σ below the mean, corresponding to a monthly attrition
rate of 1.374-1.28*(-.475)=1.98pp. The 90th percentile corresponds to a monthly attrition rate of 1.374+1.28*(-
.475)=0.77pp, a reduction of slightly more than 60%.
26
In another management-related example, Friebel et al. (2017) show that randomly sending letters from
the CEO highlighting the firm’s turnover problem causes grocery store managers to reduce store turnover by
1/3 for nine months. In a non-management example, Madrian (1994) analyzes the impact of having a spouse
with health insurance to study “job lock.” She finds that job lock reduces turnover by 25%. Thus, our IV
estimate from increasing MOR from the 10th to the 90th percentile has roughly twice the impact on turnover
as does one’s spouse having health insurance in the US. Surveying the literature, Manning (2011) reports
wage-turnover elasticities of 0.5-1.5.
27
In the data provided, turnover is marked as voluntary, involuntary, or missing. We don’t use missing
turnover events in Panels B and C (missing events are included in Panel A), but one can also classify the
missing data fields as voluntary turnover events.

21
look separately at “high” and “low” productivity workers. To classify workers as high or low
productivity, we residualize workers’ subjective performance scores on the controls in Table 3
and then regress the residuals on worker fixed effects. Fixed effects that are above the median
are classified as high-productivity workers and those below as low-productivity workers.
Appendix Table C8 analyzes turnover separately for high- and low-productivity em-
ployees. The IV coefficients are large and significant in both cases. As has been found in
many studies, there is strong selection on productivity in turnover (e.g., Hoffman and Burks,
2017), with high-productivity workers having a substantially lower base probability of attri-
tion. While the IV coefficient is larger in absolute magnitude for low-productivity workers,
it is very similar in percentage terms for high-productivity workers. Thus, it is not the case
that high-MOR managers are simply preventing low-productivity workers from leaving.
Beyond leading people to exit the firm, bad people management skills could also produce
other types of “exits.” Instead of quitting the firm, a worker may simply demand that they
be moved to a new manager. Appendix Table C9 repeats our analysis using whether a worker
changes manager as the dependent variable (instead of attrition). The IV estimate implies
that moving from a manager in the 10th percentile of MOR to one in the 90th percentile of
MOR is associated with a reduction in the probability of switching managers by 45%.
Figure 2 shows binned scatter plots for the reduced form regressions, showing a clear
negative relationship.

4.2 Research Design Based on New Workers Joining the Firm

Repeating our IV analysis while using only new workers who join the firm in the second
period, Table 4 also finds a strong negative relation between a worker’s manager’s MOR and
turnover. This “joiners” analysis has several advantages relative to our baseline analysis and
allows us to address a couple of concerns. First, in the joiners analysis, the survey responses of
the workers under analysis do not influence the instrument (i.e., the MOR that their current
manager received during period 1) because they are new to the firm. For example, one concern

22
could be that employees who have a cheery personality might be both more likely to rate their
manager highly and less likely to quit, i.e., that a worker’s cheeriness could lead to a positive
correlation between εt and u−τ . However, that issue is avoided here because MOR is measured
without influence by the cheery worker.28
Second, our analysis reduces concerns about assignment bias. When an employee joins
a very large firm, they are unlikely to have substantial information about differences in people
management skills across managers that would enable them to sort into managers. Further-
more, the firm seems unlikely to have substantial information about the new worker separate
from the manager who was involved in hiring the new worker.29 This reduces the concern that
εt is correlated with m separate from the possible role of better managers in selecting better
employees for their team.
Our sample size here is 8% of that in Table 3.

Results. Overall, while we have much less statistical power here than for our full sample,
the relationship between MOR and attrition in Table 4 is qualitatively similar to that in the
baseline analysis. Looking at all attrition in Panel A, the IV coefficient of −0.550 is slightly
larger in magnitude to that in Panel A of 3, though it just misses statistical significance at 10%
due to a much larger standard error (reflecting the smaller sample for the joiners analysis).
The coefficient implies that a 1σ increase in MOR corresponds to a 0.55pp (38%) reduction
in monthly turnover. We also see broadly similar results for quits, fires, regretted quits, and
non-regretted quits, with particularly strong results for quits (especially regretted quits), for
which the IV estimate is statistically significant.
28
The joiners analysis also eliminates concern about events in period 1 that would lead workers to rate their
managers more highly, as well as affect their quitting in period 2. For example, if an employee got to work on
a very enriching project in period 1 that improved their general human capital, this could lead them to highly
rate their manager in period 1, as well as to be less likely to quit in period 2. It is important to note, however,
that our new joiners analysis does not help with persistent shocks, e.g., a very successful or exciting project
that would make current employees rate a manager highly in period 1, as well as make new workers less likely
to quit. The joiners analysis is useful, however, if people differ in their opinions about whether a project is
exciting, as the joiners analysis looks at different people in period 2 compared to the raters in period 1.
29
Lazear et al. (2015) use new joiners as a research design for estimating manager fixed effects. Our argument
that new employees have limited private information, and that the firm has limited information (separate from
the manager), closely follows a similar argument in Lazear et al. (2015).

23
4.3 Research Design Based on Workers Joining the Firm or Chang-

ing Managers in the Second Period

While the joiners analysis has several clear benefits, it also has limitations. First, it is based
only on new workers, so there are potential concerns about external validity (e.g., could there
be something special about the impact of people management skills on new workers?). Second,
the sample size is small relative to our full sample. Third, it is difficult to do certain statistical
tests regarding assignment bias. In this section, instead of just analyzing workers joining the
firm in the second period, we additionally add instances of people switching managers during
the second period. This analysis addresses these three limitations, and we continue to find a
strong negative relation between a worker’s manager’s MOR and various turnover outcomes.
Manager switches occur for many reasons at the firm we study, such as new projects,
manager turnover, and promotions, as well as several re-organizations (“re-orgs”) that oc-
curred for exogenous reasons.30
Our analysis here is useful for multiple reasons. First, the sample is broader than
only new joiners. Second, we can do a test for non-random sorting of existing employees to
managers, following Rothstein (2010, 2017)—we happen to find little evidence of systematic
sorting of better existing employees to better people managers.31 Third, we can make “event
study” graphs analyzing how impacts of MOR on turnover vary with how long a person
has been with a manager; such a graph would be harder to interpret in the pure joiners
design, where time since manager is collinear with tenure. Fourth, analyzing what happens to
employees after changes in manager is generally useful for reducing concern about assignment
bias. While matching of managers and employees is not random, one might believe that
matching occurring for reasons such as manager turnover or re-orgs might reflect less active
30
To preserve firm confidentiality, we cannot provide detailed accounts of the re-orgs, but they were events
driven by external conditions or business conditions that affected many parts of the organization at once, and
therefore caused changes in who was managing whom.
31
As discussed further in Appendix A.5, we test whether the MOR of one’s future manager predicts employee
outcomes prior to a manager switch. This “Rothstein test” is not possible for new joiners, as employee outcomes
are unobservable prior to joining the firm. Thus, the test does not rule out that better people managers may
be adept at hiring better new employees for their teams.

24
involvement or deliberate matching of the firm, particularly when there are many individuals
being moved at the same time. It is presumably harder for a firm to do sophisticated matching
when it has to make many personnel changes in a short amount of time. Fifth, as in the
joiners analysis, the workers in this design do not influence the instrument—if I switch to a
new manager in T2 , I will have played no role in that new manager’s scores in the T1 survey.
The sample size here is about one quarter the size of that in our baseline analyses.
When workers switch manager, 4% of the time this is accompanied by a promotion in the
same month, 4% of the time they experience a change in job function, and 8% of the time
they experience a change in business unit. Thus, most changes in manager are not from
promotions, or from changes in job function or business unit.

Results. Table 5 reproduces Table 3 while analyzing both T2 joiners (as in Table 4)
and workers switching managers in T2 . In Panel A, the IV coefficient for overall attrition is
-0.332 (corresponding to a 22% drop in attrition per 1σ in MOR). This is a bit smaller than
our baseline attrition estimate and misses conventional statistical significance (p = 0.16).
In contrast, the coefficient for quitting (panel B) is slightly larger than the one in Table 3.
This drop is driven by a statistically significant drop among regretted quits. Interestingly,
the MOR coefficient for non-regretted quits is positive (though not statistically significant);
this is consistent with the idea that people management skills may play a particular rule in
reducing “bad attrition” (relative to their role in reducing “good attrition”).
As discussed in Appendix A.4, the results here are broadly similar when only using
workers switching managers.

Time path of people management impacts. Figure 3 takes the IV regressions in

Table 5, but interacts MOR with the quarter since receiving a new manager. As seen in panel
(a), when workers receive a new manager, they are less likely to attrite when the manager
has high MOR compared to when the manager has low MOR. While negative, however, the
relation between MOR and turnover in the first six months after the manager change is fairly
small in magnitude. Rather, the turnover benefit builds gradually, with much of the reduction

25
in turnover only occurring in quarters 2 and 3 after a manager change (i.e., months 7-12 since
the manager change). In the other panels, the impact of MOR also tends to grow in magnitude
with manager tenure, particularly for regretted quits.
The time path of results in Figure 3 seems consistent with a causal impact of people
management on attrition.32 The impact of a good manager may not be felt immediately after
they become a worker’s manager. Rather, it may take some time for a worker to get to know
and be affected by their manager. If the results in Figure 3 were driven by assignment bias
(e.g., the firm decides to match unobservedly better workers with better managers), one might
imagine that turnover impacts would be seen immediately instead of growing over time. The
time path also seems broadly consistent with people management playing a role in motivating
workers, as opposed to simply selecting better workers.

Rothstein test. A concern with the switchers analysis is that the firm may be matching
unobservedly high-quality managers and workers together. To test for this, we can examine
whether the MOR of an employee’s future manager predicts employee non-attrition outcomes
in the current period, following Rothstein (2010, 2017). As detailed further in Appendix A.5,
we implement an IV procedure where we instrument the future manager’s MOR during period
2 using the future manager’s MOR during period 1. Note that we cannot use attrition for the
test because workers who will experience a new manager in the future do not attrite before
they experience the new manager.
Table 6 examines the relation between a future manager’s MOR and key non-attrition
outcomes (subjective performance, salary, salary increases, and promotion propensity), as
well as with three other important worker characteristics, namely, the employee engagement,
restricted stock units granted to an employee, and whether the firm has designated a person
as a “key individual” whom they strongly want to retain. Stock grants and the key individual
variable are discussed further in Section 6 when we discuss rewards for managers.
32
One concern with this interpretation would be if the risk of attrition was changing by quarter to offset
any differences. Thus, we also created Figure 3 looking at the ratio of coefficient estimates to the mean of the
dependent variable, and obtained similar conclusions. In panel (a), the attrition rates by quarter of manager
tenure are 1.34, 1.61, 1.63, and 1.58.

26
 In the IV analyses in Panel B, we see little evidence that better people managers receive
teams with observedly better characteristics. Of the 7 coefficients, 1 is significant at 10%, close
to what might be expected from random chance. Of course, while the test suggests that better
people managers are not sorted with better employees in terms of these observables, we cannot
totally rule out that there would be sorting based on unobservables. Still, the test suggests
that assignment bias (from the firm sorting strong people managers with unobservedly better
employees) is likely limited.33

4.4 Manager Moves across Locations or Job Functions

Our third research design exploits managers moving across locations or job function. In line
with Chetty et al. (2014), we collapse our data to the location-job function-period level (e.g.,
engineers at Location ABC in period 1), and examine the relation between average MOR and
average attrition using the collapsed data. This serves two key purposes. First, it provides
further evidence (consistent with our previous tests) that assignment bias does not drive our
findings. This is because by aggregating, we focus on differences in MOR at an aggregate
level, as opposed to MOR differences within location-job function.
Second, it helps address concern about asynchronously correlated measurement error
from persistent unobservables. In our earlier joiners analysis, a persistent unobservable of
a good project could lead to employees rating their manager favorably in period 1, as well
as making new employees less likely to attrite in period 2. However, by aggregating up to
the location-job function-period level, we no longer exploit variation from some engineers at a
location working on a good project and some engineers at a location working on a bad project.

Implementation. Let Ql,f,τ,τ 0 be the (employee-month-weighted) mean normalized

MOR of managers at location l in job function f during period τ , and for which we use the
33
The test is also consistent with better people managers not selecting better existing employees for their
teams. A statistical zero in the Rothstein test could conceivably also result from negative assignment bias (in
terms of the firm’s role) and positive selection of employees by managers given the firm’s policies. Overall, we
view the Rothstein test as less important for the validity of our results (compared to other settings) given that
managers selecting better workers for their team should be viewed as a mechanism of people management.

27
measurement of the managers’ MOR taken during period τ 0 . Let yl,f,τ be the mean quit rate
of employees at location l in job function f during period τ . We estimate:

yl,f,τ = bQl,f,τ,τ 0 + δl + δf + δτ + θl,f,τ (6)

where δl , δf , δτ are location fixed effects, job function fixed effects, and period fixed effects,
respectively; and θl,f,τ is the error term. Following Chetty et al. (2014), we weight observations
by the number of employee-months per location-job function-period cell. While τ will vary
based on the cell, all cells will use the same τ 0 . That is, we measure all managers using the
same survey wave to help ensure that differences across cells reflect differences in manager
quality as opposed to different measurements. Further details on implementation are provided
in Appendix A.6.
Similar to Chetty et al. (2014), our key identifying assumption is that changes in average
location-function people management skills are uncorrelated with average location-function
unobserved determinants of attrition, conditional on controls.34 To control for possible changes
in worker quality over the two periods (due to worker sorting or workers moving with their
managers), we include controls for average location-function worker characteristics. While the
key identifying assumption is difficult to test, we are not aware of efforts by the firm (outside
of autonomous decisions by workers and managers) to optimally sort workers and managers
over time across location-job functions based on unobservables.35
In our sample, 6% of managers experience a change in location, 6% of managers experi-
ence a change in job function, and 11% of managers experience a change in either location or
job function. In the month of a location change, the promotion rate is 2.5%, whereas in the
month of a job function change, the promotion rate is 26.5%. Thus, we suspect that many
(though certainly not all) of the job function changes seem to occur from promotions, whereas
this does not seem to be the case for location changes. We suspect that the location changes
34
For even greater control, one might wish to control for location-function fixed effects instead of location
fixed effects and function fixed effects. However, we do not have enough power to do so, as doing so leads to
very large standard errors. Instead, we control for a rich set of location-function worker characteristics.
35
Now, it remains quite possible or likely that good people managers may start selecting or attracting better
workers to a location upon moving there. As discussed above, such selection may be a key mechanism by
which people management skills matter.

28
involve a combination of business reasons (e.g., moving to a nearby location because of some
business need) and family reasons.

Results. We obtain the same broad conclusion that people management skills reduces
attrition, even though we are exploiting a quite different source of variation in MOR than in
our baseline analyses. Columns 1-2 of Table 7 show OLS results, one measuring all managers
in the sample using their wave 1 score, and the other measuring all managers using their wave
2 score. To account for possible attenuation bias due to measurement error, columns 3-4 show
IV results, where we instrument the mean MOR in the location-job function-period cell using
mean MOR in the other period for that location-job function.
As in our main results in Table 3, the IV estimates here are larger in magnitude than
the OLS estimates. Focusing first on the overall attrition results in Panel A, the IV estimates
imply that a 1σ increase in a manager’s MOR decreases employee attrition by 0.6-0.7pp per
month, which is a bit larger in magnitude than our benchmark estimate in Panel A of Table
3, though our IV confidence intervals here clearly overlap with that in Panel A of Table 3.
Outside of Panel A, we have less power, but observe qualitatively similar results to before.
Appendix Table C11 shows that the results here are robust to (and actually a bit stronger
when) restricting to location-job functions that are in the data for both periods.

4.5 Additional Threats to Identification and Robustness

Adding Richer Controls. A concern for the results is the possibility of a persistent unob-
servable that is not fully addressed by the above research designs (e.g., people always rate a
manager well because that manager is overseeing a good project, and the manager continues
working on the good project when he/she moves locations or job functions). A way to proxy
for such unobservables is to add further controls.
Appendix Tables C12-C15 shows that our attrition estimates are quite similar when
adding additional controls. For the 4 sets of results (baseline + 3 ID strategies), we gradually
add additional controls, including two-way interactions between business unit, job function,

29
and salary grade, as well as current month dummies. For example, instead of just having
dummies for being an engineer and being at a particular salary grade, we add dummies for
being an engineer of a particular salary grade. To formally assess coefficient sensitivity, we
perform the Oster (2017) test. Given our IV set-up, we analyze the reduced form. The results
consistently imply that selection on unobservables would need to be many times larger than
selection on observables (details are in Appendix A.7).

Functional Form of the Regressor. Our results use normalized MOR. To check that
our results are not driven by the functional form of the regressor or by outliers, we re-ran the
analysis in Panel A of Table 3 while grouping MOR in percentiles or 5 quintiles, and obtained
qualitatively similar results. This is unsurprising given the fairly linear relationship in the
reduced form in Figure 2. Appendix Figures C2 and C3 repeat our results using the research
designs in Sections 4.2-4.3.

Heterogeneity analysis. Within the large firm we study, we can also examine het-
erogeneity in the IV estimates by geography, occupation, and hierarchy. Appendix Tables
C16-C18 show that the negative IV relation between MOR and turnover is qualitatively ro-
bust across geography, occupation, and hierarchy. For brevity, results are discussed in detail
in Appendix A.8.

VA Approach. We can also perform a “value-added” analysis of managers, similar to

Lazear et al. (2015) and Hoffman et al. (2018). (We use quotation marks because we analyze
attrition instead of productivity.) By doing so, we can compare the spread of manager VA
with the importance of MOR for attrition. In line with our focus on attrition, we estimate:

yit = α + γj + Xit δ + it (7)

where yit is a dummy for whether person i attrites in month t; γj is a manager effect; and Xit
are various controls. As discussed in Lazear et al. (2015) and Hoffman et al. (2018), as well
as in work on teacher VA, an important issue is accounting for sampling variation. That is,
if manager fixed effects are estimated using a finite number of observations per manager, our

30
estimate of the standard deviation of manager fixed effects may be biased upward. We address
this using two approaches. First, we present standard deviations weighted by the number of
observations per manager (“one sample” approach). Second, following Silver (2016), we split
the data in two and estimate (7) for two separate samples. Assuming that the sampling error
is uncorrelated across samples and uncorrelated with underlying value-add, the covariance of
the estimated manager VA across the two samples is equal to the variance of underlying VA.36
We do this either randomly splitting employee-months into two samples or splitting by period.
Appendix Table C19 shows that there is significant variation in manager effects for
the outcome of employee attrition. In the split sample approach, we find that the standard
deviation of manager effects for attrition is 0.67 when splitting randomly and by period. Thus,
the consequence of improving manager VA in attrition by 1σ (0.67pp per month) is about 40%
larger than the impact of improving underlying people management skills by 1σ in our baseline
estimate in Panel A of 3 (0.48pp per month). Appendix A.9 discusses VA further.

5 Manager Quality and Non-attrition Outcomes

This section shows that there is no consistent evidence MOR improves non-attrition outcomes.
There is a small statistically significant relation between MOR and subjective performance,
and a fairly precise no relation of MOR with either salary increases or promotion. The small
positive relation for subjective performance is not robust to our research designs.

Employee subjective performance. Columns 1-2 of Table 8 show that MOR appears
to have only a modest positive relation to employee performance as measured by subjective
performance reviews. On the left-hand side, we use an employee’s subjective performance
review on a 1-5 scale, which we then normalize. Column 1 of Table 8 presents a baseline
estimate without employee fixed effects. A 1σ increase in MOR is associated with a 0.04σ
increase in employee subjective performance under OLS, as well as a 0.10σ increase under IV.
36
Let γ be the underlying VA for a manager. Let γˆ1 = γ + u1 and γˆ2 = γ + u2 be estimated VA in the two
split samples, where u1 and u2 are errors. Note that cov(γˆ1 , γˆ2 ) = var(γ)+cov(γ, u1 )+cov(γ, u2 )+cov(u1 , u2 ),
which equals var(γ) under the stated assumptions. This derivation also appears in Silver (2016).

31
As for the earlier attrition results, OLS is likely biased downward due to attenuation bias.
In column 2, we add employee fixed effects. It is not clear a priori whether the results
with or without fixed effects should be preferred. The results without employee fixed effects
examine the relationship between MOR and employee outcomes inclusive of managers possibly
being able to select better employees. Results with employee fixed effects tell us how MOR
relates to various outcomes within an employee, which may be useful to know if some managers
happen to receive better or worse employees as a result of luck or other factors unrelated to
their managerial quality. We therefore will often present results with and without employee
fixed effects. In column 2, when employee fixed effects are included, the relationship between
MOR and subjective performance shrinks toward 0. The IV coefficient of 0.032 seems small in
economic magnitude, its magnitude implying that moving from the 10th to 90th percentile of
managers would only improve subjective performance by 0.08σ. Column 2 suggests that the
estimate in column 1 reflects some aspect of how managers and workers are sorted together
(such as better managers hiring better workers).

Employee salary increases. In Columns 3-4 of Table 8, the outcome is the increase
in salary 12 months from now relative to the present. That is, for an employee in May Y1 ,
the outcome variable is log(salary) in May Y2 minus log(salary) in May Y1 . In column 3, a 1σ
increase in MOR is associated with roughly a 0.12% increase in employee salary in OLS and
a 0.06% increase in IV, both statistically insignificant. The average salary increase per year
in our data is confidential, but is between 4% and 8%, so these impacts seem small compared
to that. With 95% confidence, we rule out coefficients greater than 0.28-0.47 (depending on
OLS or IV), i.e., we rule out that a 1σ increase in MOR predicts a salary increase more than
0.28%-0.47%. The coefficients are broadly similar when employee fixed effects are added.

Employee promotions. Columns 5-6 examine the relationship between MOR and
employee promotions. In column 5 of Panel A, we see an insignificant positive relation in
the OLS, but it turns negative in the IV in Panel B. In the IV in column 5, the top of the
95% confidence interval is 0.24. Given the average monthly promotion rate at the firm of

32
between 1.5% and 2%, this means we can rule out that a 1σ increase in MOR would increase
the promotion probability by about 12%-16%. Thus, while we cannot rule out small impacts,
we can rule out sizable ones. Recall we find 1σ decreases attrition by 28%, so the top of the
confidence interval for promotion is about half as large.

5.1 Additional Remarks on Non-attrition Outcomes Results

Research designs. Appendix Tables C20 and C21 repeat our joiners and workers changing
managers research designs for the non-attrition outcomes. Compared to attrition, the results
on the non-attrition outcomes are much less robust across the different research designs. Thus,
while the different designs provide strong evidence that better people management reduces
attrition, they fail to do so for other outcomes.

Additional outcome. Beyond the above outcomes, another outcome we can evalu-
ate is employee engagement, which is obtained from the same survey as the manager scores.
Appendix Table C22 shows some evidence that MOR may predict increased employee satis-
faction, though the coefficients are small in magnitude and vary across specification. A 1σ
in people management skills predicts a 0.07σ increase in employee engagement in IV without
worker fixed effects, but only a statistically insignificant 0.02σ increase in IV with fixed effects.

Differential Attrition. In Section 4, we provided evidence that higher MOR lowers

employee attrition. However, such differential attrition could potentially bias estimation of
the relation between MOR and non-attrition outcomes. For example, if a very good manager
successfully retains all of their employees (both the stars and the mediocre ones), this might
lead to the manager getting lower average achievement on an employee outcome variable than
if the mediocre employees left. To address this concern, we repeated our analysis in Table
8 while restricting to employees who never attrite. Such a “balanced panel” analysis (e.g.,
Brown et al., 2016; Burks et al., 2015) yields qualitatively similar results to those in Table 8.

Why do we see large impacts of people management on attrition, but not

on our non-attrition outcomes? There are several possible answers. First, it could be

33
that good people management naturally matters most for attrition, which may reflect issues
such as whether an employee feels respected and motivated. The people management skills of
one’s manager may matter less for subjective performance, salary growth, or promotions, for
which technical talent and knowledge may be more important. Second (and related to the first
answer), it could reflect that some outcomes are “stickier” and harder to change. It might be
easier for a manager to affect whether someone feels engaged and motivated (thereby reducing
attrition) and harder to affect subjective performance. Such a possibility would not diminish
the importance of attrition, but would reflect that it is easier to change. Third, it could be
that certain outcomes take more time and more interaction to be affected. Ultimately, it is
hard for us to definitively distinguish these possibilities in our data. The primary contribution
of our paper is how people management skill relates to different outcomes.

6 How does the Firm Reward Good Managers?

So far, we have presented evidence that a manager’s people management skills, as measured
by MOR, reduce employee turnover. We now examine whether MOR is “rewarded” by the
firm in terms of how it evaluates, compensates, and promotes its managers. In large high-skill
firms such as the one we study, the concept of reward may be complex and multi-faceted.
Individuals can be rewarded through promotions, higher salary, or stock grants. The firm
could also respond to managers in other ways such as changing their span of control so that
better managers become responsible for managing more people.37
We estimate regressions similar to those in Section 3.1, except we include manager
rewards on the left-hand side instead of employee outcomes. For OLS, this would be:

Rj,t = bm
e j,τ (t) + θj,t (8)
37
Another outcome to examine is whether a manager attrites. Our IV analysis is not ideally suited for
analyzing manager attrition because it requires observing a manager score for a manager in both periods.
This caveat aside, if one performs the analysis in Table 9 using manager attrition as an outcome, one finds a
statistically significant negative relation between a manager’s MOR and the manager’s turnover, i.e., managers
with better people management skills are less likely to attrite.

34
where Rj,t is a reward (e.g., subjective performance score) achieved by manager j in month t.
We include the same controls as for our analysis of worker outcomes.

6.1 Evidence on Manager Rewards

Manager Subjective Performance. Subjective performance is a critical measure of “re-

ward” at our firm, as the subjective performance score is a critical determinant (though not
an absolutely deterministic one) of the financial rewards that a person receives. As seen in
column 1 of Table 9, a 1σ increase in MOR is associated with a 0.06σ increase in subjective
performance in OLS, but a 0.40σ increase in subjective performance in IV. The IV coefficient
is substantial, both statistically and in economic magnitude. As in our main results on how
MOR affects employee attrition, we suspect that OLS is subject to attenuation bias from
measurement error.

Promotions. Table 9 additionally shows that better people managers are substantially
more likely to get promoted, with a 1σ increase in MOR predicting a 0.7pp increase in pro-
motion probability each month. Given the average monthly promotion rate of 1.5%-2%, this
coefficient is economically substantial. The result is consistent with the firm wishing to pro-
mote good people managers to higher levels of the organization where they may have greater
impact, i.e., a selection story. It is also consistent with an incentive story, where the firm
encourages good people management skills by promoting those who exhibit them (Lazear and
Rosen, 1981).38

Compensation. There no significant relationship between MOR and log salary. How-
ever, there is a statistically significant positive relationship between MOR and log salary
growth. A 1σ increase in MOR is associated with a 1.4% larger increase in salary over a 12-
month period. As reported earlier, the average annual increase in salary is confidential, but
is between 4-8%, so this seems economically meaningful. If we re-do the result using 1-month
38
While this result is intuitive, it is far from obvious. If a manager is doing well, there is no guarantee that
they would also succeed in a new role after a promotion. Such considerations seem to be given less weight by
the firm than the selection and/or incentive rationales of promoting people with high MOR.

35
increase in log salary, there is also a significantly positive IV relationship.
Another means of compensation, particularly in high-skill firms, is restricted stock
grants, which are given to reward and retain valued employees. Table 9 shows that there
is no relation between MOR and the level of an employee’s stock grant holdings, or between
MOR and disbursement of new stock grants.
Thus, higher people management skills predict high salary growth, but not the level of
salary or stock grants. Overall, it seems that there is no consistent relationship between MOR
and compensation, though there is also some evidence for a positive relationship.

Span of Control. In models of optimal span of control such as Lucas (1978) and
Garicano (2000), firms optimally assign better managers to manage larger teams. Empirically,
we examine whether managers who achieve higher MOR become more likely to manage larger
teams. Column 7 shows that an increase in MOR by 1σ predicts an increase in span of control
by 0.3 individuals, but it is not statistically significant (though the standard error of 0.3 also
seems relatively large).

Key individual designation. Individuals at the firm who are believed to be especially
important can be designated by the firm as “key individuals.” The data show that better
people managers are more likely to be designated “key individuals” (with a 1σ increase in
MOR predicts a 2.1pp increase in the probability of being designed a key individual), though
the relation is not statistically significant.

6.2 Threats to Identification and Discussion

Adding richer controls. A concern for our rewards results in whether they could reflect
some unobserved variable. For example, if there was a persistent unobservable that affected
the way employees ranked their manager as well as how a manager was rewarded (e.g., a good
project), this could be a violation of the exclusion restriction.
Similar to as in Section 4.5, Appendix Table C23 presents results on the statistically
significant reward variables (subjective performance, promotions, and salary increases) as

36
further controls are added. The key IV coefficient is fairly stable across specifications, with
the limited selection on observables suggesting that selection on unobservables is also likely
small (Oster, 2017).

Pay for past performance. An issue in considering manager rewards is whether pay
for past performance could constitute a violation of the exclusion restriction. To address
this, we redo our compensation results while restricting attention to compensation in the
first period (so that the instrumental variable is a manager’s people management score in the
second period). As seen in Appendix Table C24, the results are broadly qualitatively similar.39

Including manager VA as a regressor. One question is whether the results are

robust to controlling for a manager’s overall skill at reducing attrition. That is, conditional
on a manager’s ability to reduce attrition, are managers with greater underlying people man-
agement skills still rewarded? To address this question, we repeat Table 9 while including a
manager’s attrition VA fixed effect on the right-hand side. We normalize the fixed effects and
multiple by -1 to create a manager fixed effect in terms of retention. To account for sampling
error in manager VA, we use a split sample IV approach (e.g., Frederiksen et al., 2017); as in
Section 4.5, we estimate manager fixed effects separately after randomly splitting the data in
two. We use one fixed effect as the endogenous regressor and one as the instrument.
Appendix Table C25 shows that the relation of MOR to rewards is mostly similar when
controlling for manager VA. Appendix A.10 provides further discussion.

Discussion. Broadly speaking, Table 9 shows that better people managers do receive
significant rewards from the firm in some important dimensions. Even though the firm has a
strong engineering culture and values technical skills, it still appears to reward good people
management. In fact, if one redoes the analysis by occupation, there is a stronger relation
between MOR and a manager’s subjective performance score for engineers (IV coef=0.99,
39
For subjective performance, the IV coefficient is 0.20, which is statistically significant at 5%, and a bit
smaller than in Table 9, though the 95% confidence intervals overlap. For promotions, the IV coefficient of 0.94
is a bit higher than in Table 9, but is no longer statistically significant, reflecting a large standard error. For
compensation, the coefficient on log salary growth remains positive but loses statistical significance, whereas
the coefficient on log change in stock grants increases (but is not statistically significant).

37
se=0.32), relative to the overall population.
One important caveat regarding the reward results is that we only observe a relatively
short panel. In the longer-run, it could be that better people managers are rewarded to a
greater (or lesser) extent than observed here.

7 Conclusion

Managers are at the heart of organizations, but measuring what managers do is challenging.
An approach taken in studying CEOs and managers in lower-skill firms has been to calculate
a manager’s VA using performance metrics. However, such an approach may be difficult in
knowledge-based firms and other firm contexts where objectively measuring productivity is
challenging. We pursue an alternative approach using employee surveys. Employee surveys
also enable us to address a different and complementary question to VA work. VA papers an-
swer: how much do managers matter? We answer: how much does one particular set of skills,
namely, people management skills, or interpersonal skills for dealing with one’s subordinates,
matter? Employee surveys about managers are used by many firms, but we have little hard
evidence on the importance of people management skills.
In our baseline IV results, we find a strong negative relationship between people man-
agement skills and employee attrition, a critical outcome in high-skill firms. A causal in-
terpretation is strengthened using several complementary research designs. However, people
management skills do not seem to consistently improve non-attrition employee outcomes.
Managers with better people management skills receive higher performance ratings and are
more likely to be promoted, consistent with the firm attaching significant value to managers’
impacts on employee turnover.
Although our conclusions are specific to one firm, our main findings on the importance
of people management skills are robust across different geographies, occupations, hierarchy
levels within the large, multinational firm we study. This strengthens the case for the external
validity of our results (which is frequently a challenge in personnel economics), and suggests

38
that our conclusions may hold in other contexts. While statistical power is limited, there is
some intriguing evidence that people management skills seem to be particularly important in
rich countries and at high levels of the organization. Because the surveys we study are collected
by many firms, future work should examine how the importance of people management skills
vary by firm and context.
By evidencing the importance of good people management, our paper highlights an as-
pect of managers that differs from that emphasized by most theories of managers. Moral
hazard models (e.g., Holmstrom, 1979) emphasize the importance of managers for helping ad-
dress employee moral hazard (e.g., by monitoring) or by making resource allocation decisions.
Knowledged-based models, beginning with Garicano (2000), emphasize the role of managers
in problem-solving, i.e., a good manager can solve more complex problems than their direct
reports. We see an open role for theory in constructing models of people management skills,
e.g., models where managerial skills affect relationships between employees and managers.
One direction related to our paper is the growing literature on social skills, i.e., skills that
are primarily interpersonal in nature. Scholars have shown that social skills play a critical role
in determining wages and occupations, and that social skills command a rising return in the
labor market (e.g., Borghans et al., 2014; Deming, 2017).40 Our work provides novel evidence
on the importance of social skills in management, an area where they have received much less
economic attention than others, but where new evidence is starting to emerge.41
Though our results indicate that people management skills matter to a significant degree,
the precise mechanism by which they matter remains an important area of research. Theories
of managerial attention (e.g., Dessein et al., 2016) emphasize the importance of attention as a
limited resource in determining productivity across managers. We hope to be able to examine
such theories in future work.
40
Social skills are generally thought of as one component of “soft skills;” see Heckman and Kautz (2012) for
a general discussion on soft skills.
41
Kuhn and Weinberger (2005) show that men who occupied high school leadership positions earn more
years later. Lazear (2012) shows that successful business leaders tend to be generalists instead of specialists.
Schoar (2016) shows that a randomized intervention in Cambodian garment factories aimed at improving
supervisors’ communication skills and treatment of workers improved productivity.

39
References
Aghion, Philippe, Benjamin F. Jones, and Charles I. Jones, “Artificial Intelligence
and Economic Growth,” WP 23928, National Bureau of Economic Research October 2017.

Ashenfelter, Orley and Alan Krueger, “Estimates of the Economic Return to Schooling
from a New Sample of Twins,” American Economic Review, 1994, 84 (5), 1157–1173.

and Cecilia Rouse, “Income, Schooling, and Ability: Evidence from a New Sample of
Identical Twins,” Quarterly Journal of Economics, 1998, 113 (1), 253–284.

Atkins, Paul WB and Robert E Wood, “Self-versus Others’ Ratings as Predictors of

Assessment Center Ratings: Validation Evidence for 360-degree Feedback Programs,” Per-
sonnel Psychology, 2002, 55 (4), 871–904.

Baker, George P., Michael Gibbs, and Bengt Holmstrom, “The Wage Policy of a
Firm,” Quarterly Journal of Economics, 1994, 109 (4), pp. 921–955.

, Robert Gibbons, and Kevin J. Murphy, “Subjective Performance Measures in Op-

timal Incentive Contracts,” Quarterly Journal of Economics, 1994, 109 (4), 1125–1156.

Balasubramanian, Natarajan, Jin Woo Chang, Mariko Sakakibara, Jagadeesh

Sivadasan, and Evan P. Starr, “Locked In? The Enforceability of Covenants Not to
Compete and the Careers of High-Tech Workers,” 2017. Working Paper, U. Michigan.

Bandiera, Oriana, Stephen Hansen, Andrea Prat, and Raffaella Sadun, “CEO Be-
havior and Firm Performance,” Working Paper 23248, National Bureau of Economic Re-
search March 2017.

Bartel, Ann P., Brianna Cardiff-Hicks, and Kathryn Shaw, “Incentives for Lawyers:
Moving Away from Eat What You Kill,” ILR Review, 2017, 70 (2), 336–358.

Beleche, Trinidad, David Fairris, and Mindy Marks, “Do course evaluations truly
reflect student learning? Evidence from an objectively graded post-test,” Economics of
Education Review, 2012, 31 (5), 709–719.

Bender, Stefan, Nicholas Bloom, David Card, John Van Reenen, and Stefanie
Wolter, “Management Practices, Workforce Selection, and Productivity,” Journal of Labor
Economics, 2018, 36 (S1), S371–S409.

Bertrand, Marianne and Antoinette Schoar, “Managing with Style: The Effect of Man-
agers on Firm Policies,” Quarterly Journal of Economics, 2003, 118 (4), 1169–1208.

Bloom, Nicholas and John Van Reenen, “Measuring and Explaining Management Prac-
tices Across Firms and Countries,” Quarterly Journal of Economics, November 2007, 122
(4), 1351–1408.

and John Van Reenen, “Human Resource Management and Productivity,” Handbook
of Labor Economics, 2011, 1, 1697–1767.

40
 , Erik Brynjolfsson, Lucia Foster, Ron S. Jarmin, Megha Patnaik, Itay Saporta-
Eksten, and John Van Reenen, “What drives differences in management?,” WP 23300,
National Bureau of Economic Research 2017.

, James Liang, John Roberts, and Zhichun Jenny Ying, “Does working from home
work? Evidence from a Chinese experiment,” Quarterly Journal of Economics, 2014, 130
(1), 165–218.

Borghans, Lex, Bas Ter Weel, and Bruce A Weinberg, “People Skills and the Labor-
market Outcomes of Underrepresented Groups,” ILR Review, 2014, 67 (2), 287–334.

Bound, John and Alan B. Krueger, “The Extent of Measurement Error in Longitudinal
Earnings Data: Do Two Wrongs Make a Right?,” Journal of Labor Economics, 1991, 9 (1),
1–24.

, Charles Brown, and Nancy Mathiowetz, “Measurement Error in Survey Data,”

Handbook of Econometrics, 2001, 5, 3705–3843.

Brown, Meta, Elizabeth Setren, and Giorgio Topa, “Do Informal Referrals Lead to
Better Matches? Evidence from a Firms Employee Referral System,” Journal of Labor
Economics, 2016, 34 (1), 161–209.

Brutus, Stéphane, Mehrdad Derayeh, Clive Fletcher, Caroline Bailey, Paula Ve-
lazquez, Kan Shi, Christina Simon, and Vladimir Labath, “Internationalization of
multi-source feedback systems: A six-country exploratory analysis of 360-degree feedback,”
The International Journal of Human Resource Management, 2006, 17 (11), 1888–1906.

Burks, Stephen V., Bo Cowgill, Mitchell Hoffman, and Michael Housman, “The
Value of Hiring through Employee Referrals,” Quarterly Journal of Economics, 2015, 130
(2), 805–839.

Carrell, Scott E. and James E. West, “Does Professor Quality Matter? Evidence from
Random Assignment of Students to Professors,” Journal of Political Economy, 06 2010, 118
(3), 409–432.

Chetty, Raj, John N. Friedman, and Jonah E. Rockoff, “Measuring the Impacts
of Teachers I: Evaluating Bias in Teacher Value-Added Estimates,” American Economic
Review, September 2014, 104 (9), 2593–2632.

Clark, Andrew E., “What Really Matters in a Job? Hedonic Measurement using Quit
Data,” Labour Economics, 2001, 8 (2), 223–242.

Deming, David J., “The Growing Importance of Social Skills in the Labor Market,” Quar-
terly Journal of Economics, 2017, 132 (4), 1593–1640.

Dessein, Wouter, Andrea Galeotti, and Tano Santos, “Rational Inattention and Or-
ganizational Focus,” American Economic Review, June 2016, 106 (6), 1522–36.

Frederiksen, Anders, “Job Satisfaction and Employee Turnover: A Firm-Level Perspec-

tive,” German Journal of Human Resource Management, 2017, 31 (2), 132–161.

41
 , Lisa B. Kahn, and Fabian Lange, “Supervisors and Performance Management Sys-
tems,” Working Paper 23351, National Bureau of Economic Research April 2017.

Friebel, Guido, Matthias Heinz, and Nick Zubanov, “Making Managers Matter,” 2017.
Working paper, Frankfurt University.

Garicano, Luis, “Hierarchies and the Organization of Knowledge in Production,” Journal

of Political Economy, 2000, 108 (5), 874–904.

Garvin, David A, Alison Berkley Wagonfeld, and Liz Kind, “Google’s Project Oxygen:
Do Managers Matter?,” 2013, Harvard Business School Case Study.

Glover, Dylan, Amanda Pallais, and William Pariente, “Discrimination as a Self-

Fulfilling Prophecy: Evidence from French Grocery Stores,” Quarterly Journal of Eco-
nomics, 2017, 132 (3), 1219–1260.

Harvard Management Update, “How Great Managers Manage People,” Harvard Business
Review, 2008, February 28.

Heckman, James J. and Tim Kautz, “Hard Evidence on Soft Skills,” Labour Economics,
2012, 19 (4), 451–464.

Hoffman, Mitchell and Stephen V. Burks, “Training Contracts, Employee Turnover, and
the Returns from Firm-sponsored General Training,” 2017. NBER Working Paper 23247.

, Matthew Bidwell, John McCarthy, and Michael Housman, “The Determinants

of Managerial Productivity around the World,” 2018. Mimeo, University of Toronto.

Holmstrom, Bengt, “Moral Hazard and Observability,” The Bell Journal of Economics,
1979, pp. 74–91.

Ichniowski, Casey, Kathryn Shaw, and Giovanna Prennushi, “The Effects of Hu-
man Resource Management Practices on Productivity: A Study of Steel Finishing Lines,”
American Economic Review, 1997, 87 (3), 291–313.

Jovanovic, Boyan, “Job Matching and the Theory of Turnover,” Journal of Political Econ-
omy, 1979, 87 (5), 972–90.

Kuhn, Peter and Catherine Weinberger, “Leadership Skills and Wages,” Journal of
Labor Economics, 2005, 23 (3), 395–436.

Kuhnen, Camelia M. and Paul Oyer, “Exploration for Human Capital: Evidence from
the MBA Labor Market,” Journal of Labor Economics, 2016, 34 (S2), S255–S286.

Lazear, Edward P., “Leadership: A Personnel Economics Approach,” Labour Economics,

2012, 19 (1), 92–101.

and Sherwin Rosen, “Rank-Order Tournaments as Optimum Labor Contracts,” Journal

of Political Economy, October 1981, 89 (5), 841–64.

42
 , Kathryn Shaw, and Christopher Stanton, “The Value of Bosses,” Journal of Labor
Economics, 2015, 33 (4), 823–861.

Lucas, Robert E., “On the Size Distribution of Business Firms,” The Bell Journal of Eco-
nomics, 1978, pp. 508–523.

Madrian, Brigitte C., “Employment-Based Health Insurance and Job Mobility: Is There
Evidence of Job-Lock?,” Quarterly Journal of Economics, 1994, 109 (1), 27–54.

Manning, Alan, “Imperfect competition in the labor market,” Handbook of Labor Eco-
nomics, 2011, 4, 973–1041.

Oster, Emily, “Unobservable Selection and Coefficient Stability: Theory and Evidence,”
Journal of Business & Economic Statistics, 2017, 0 (0), 1–18.

Pfau, Bruce, Ira Kay, Kenneth M Nowack, and Jai Ghorpade, “Does 360-degree
feedback negatively affect company performance?,” HR Magazine, 2002, 47 (6), 54–59.

Pischke, Jorn-Steffen, “Lecture Notes on Measurement Error,” 2007. URL: http://econ.

lse.ac.uk/staff/spischke/ec524/Merr_new.pdf. Last visited on 2017/07/06.

Rothstein, Jesse, “Teacher Quality in Educational Production: Tracking, Decay, and Stu-
dent Achievement,” Quarterly Journal of Economics, 2010, 125 (1), 175–214.

, “Measuring the Impacts of Teachers: Comment,” American Economic Review, June 2017,
107 (6), 1656–84.

Schoar, Antoinette, “The Importance of Being Nice: Supervisory Skill Training in the
Cambodian Garment Industry,” 2016. Mimeo MIT.

Shankar, Kameshwari and Suman Ghosh, “A Theory of Worker Turnover and Knowledge
Transfer in High-Technology Industries,” Journal of Human Capital, 2013, 7 (2), 107–129.

Shaw, Kathryn and Debra Schifrin, “Royal Bank of Canada: Transforming Managers
(A),” 2015, Stanford GSB Case Study.

Silver, David, “Haste or Waste? Peer Pressure and the Distribution of Marginal Returns to
Health Care,” 2016. Mimeo, UC Berkeley.

Stoyanov, Andrey and Nikolay Zubanov, “Productivity Spillovers across Firms through
Worker Mobility,” American Economic Journal: Applied Economics, 2012, 4 (2), 168–98.

Yukl, Gary, Leadership in Organizations, 7th ed., Prentice Hall, 2010.

43
 Figure 1: Correlation of Manager Overall Rating (MOR) across Two Surveys

90
85 80
MOR in P2
75 70

Coefficient = 0.371
(0.031)
65

40 60 80 100
MOR in P1

Notes: This figure presents a binned scatter plot of MOR in period 2 on MOR in period 1 (using
“binscatter” in Stata), with no control variables. An observation is a manager. In the lower-right of the
figure, we list the regression coefficient (with a robust standard error in parentheses) for a manager-level
regression of MOR in period 2 on MOR in period 1. See Appendix Figure C1 for a similar figure, but made
separately for each of the six manager survey questions.

44
Figure 2: Reduced Form Binned Scatter Plots: Regressing Attrition Variables on Current
Manager MOR in Other Period
1.8

1
.9
1.6

.8
Coef

Coef
1.4

.7
1.2

.6
Coefficient = -0.154 Coefficient = -0.090
(0.032) (0.023)
1

.5
-2 -1 0 1 2 -2 -1 0 1 2
MOR in Other Period MOR in Other Period

(a) Attrition (b) Quits

.8
.5

.7
.4
Coef

Coef
.6
.3

.5

Coefficient = -0.061 Coefficient = -0.075

.4

(0.015) (0.020)
.2

-2 -1 0 1 2 -2 -1 0 1 2
MOR in Other Period MOR in Other Period

(c) Fires (d) Regretted Quits

.3

7
.25

6.5
6
Coef

Coef
.2

5.5
.15

5
.1

Coefficient = -0.015 Coefficient = -0.443

(0.010) (0.174)
4.5

-2 -1 0 1 2 -2 -1 0 1 2
MOR in Other Period MOR in Other Period

(e) Non-regretted Quits (f) Worker changes manager

Notes: This figure presents binned scatter plots corresponding to the reduced form regressions in Table 3.
Controls are the same as in Table 3.

45
Figure 3: Impacts of MOR on Attrition Outcomes by Quarter Since Getting New Manager

.5
.5
0

0
IV coefficient estimate

IV coefficient estimate
-.5

-.5
-1

-1
-1.5

-1.5
-2

0 1 2 3 0 1 2 3
Quarter relative to manager change Quarter relative to manager change

(a) Attrition (b) Quits

.5

0
IV coefficient estimate

IV coefficient estimate
0

-.5
-.5

-1
-1

-1.5

0 1 2 3 0 1 2 3
Quarter relative to manager change Quarter relative to manager change

(c) Fires (d) Regretted Quits

.6

2
.4

0
IV coefficient estimate

IV coefficient estimate
-2
.2

-4
0
-.2

-6

0 1 2 3 0 1 2 3
Quarter relative to manager change Quarter relative to manager change

(e) Non-regretted Quits (f) Worker changes manager

Notes: Dotted line shows 90% confidence interval on coefficients, with standard errors clustered by manager.
This figure comes from an IV regression similar to that in Table 5, with one main difference. The difference
is that instead of using MOR, we use MOR interacted with quarters since getting a new manager. “Quarter
0” includes the month during which a worker gets a new manager, followed by the two months after (i.e.,
months 2 and 3). “Quarter 3” includes months 10, 11, and 12. Beyond quarters 0-3 shown here, we also
include a single dummy for being in quarters 4 or 5 (this is a small bin, including about 5% of observations
in the analysis, whereas the other bins each include about 10% or more). Both current period MOR (the
regressor) and other period MOR (the instrument) are 46interacted with quarters since getting a new manager.
 Table 1: Summary Statistics

Panel A: Overall numbers

Share of records, employee in US 0.70
Share of records from managers 0.21
Share of records for engineers 0.36
Co-located with manager 0.81
Same function as manager 0.86
Average manager span (employees/mgr) 9.35
Managers per employee in the sample 1.39
Managers per employee (weighted by tenure) 1.52

Panel B: Several outcomes and

regressors of interest
Variable: mean sd min max
Attrition probability (monthly) x100 1.37 11.64 0 100
Quit probability (monthly) x100 0.79 8.86 0 100
Fire probability (monthly) x100 0.29 5.39 0 100
Regretted quit prob (monthly) x100 0.62 7.86 0 100
Non-regretted quit prob (monthly) x100 0.17 4.11 0 100
Subjective performance rating 3.33 0.82 1 5
Log salary Confidential
Promotion probability (monthly) Confidential
Manager overall rating (MOR) 80.86 15.27 15 100
Manager gives clear expectations 84.23 13.5 0 100
Manager provides coaching 76 17 0 100
Manager supports career dev 77.7 16.31 0 100
Manager involves people 84.27 14.36 0 100
Manager instills poz attitude 82.95 15.57 0 100
Manager is someone I trust 82.49 15.03 13 100
Notes: This table presents important summary statistics regarding our sample. The data are at the
employee-month level. The number of observations is withheld to protect firm confidentiality. We also
cannot disclose the exact time frame of the sample, but the sample is for a 27-month period between
January Y1 and March Y3 in 2011-2015. Thus, Y1 corresponds to 2011, 2012, or 2013, but we cannot disclose
which year it is. In Panel A, “Share of records, employee in US” refers to the share of employee-months in
the dataset where the employee is working at a US location. “Co-located with manager” refers to the share
of employee-months where the employee and manager are working at the same location. For further detail on
sample construction, see Appendix B. Appendix Table C1 repeats this table using the data before imposing
the restriction that MOR be non-missing for the current manager in both surveys.

47
 Table 2: Managerial Characteristics are Persistent: Predicting Manager Ratings on Different Dimensions in the Y2 Survey
using Ratings from the Y1 Survey

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

Dep. Variables: Overall Clear Coaching Career dev Involves Positive Someone
MOR expectations people attitude I trust
Characteristic in Y1 0.37*** 0.25*** 0.29*** 0.31*** 0.29*** 0.43*** 0.35***
(0.04) (0.04) (0.03) (0.03) (0.04) (0.04) (0.04)
R-squared 0.233 0.183 0.205 0.224 0.179 0.251 0.205
Notes: Robust standard errors in parentheses. An observation is a manager. Each column regresses a managerial score variable in Y2 on the same
variable in Y1 . For example, column 1 regresses a manager’s overall rating (MOR) in Y2 on a manager’s MOR in Y1 as well as control variables. The
sample is restricted to managers for whom we have manager scores for both waves of the employee surveys. We include control variables
corresponding to a manager’s first observation in the data as a manager. All regressions include controls for business unit, for work type (9
categories), dummies for year of hire (observations before 2001 are lumped in one year), salary grade dummies, and location dummies. Locations with
less than 2,000 employee-months are lumped into a separate location category, and we also include a separate dummy variable for a location being in
the US. The questions from the survey are listed in the main text in Section 2.3. * significant at 10%; ** significant at 5%; *** significant at 1%
48
 Table 3: MOR and Employee Attrition: Baseline Results

Specification: 1st Stg OLS IV Reduced Form

Panel A: Attrition
MOR in other period 0.325*** -0.154***
(0.029) (0.032)
MOR in current period -0.156*** -0.475***
(0.031) (0.103)
Mean dep. var. 1.374 1.374 1.374
F-stat on excl instrument 124.6
Panel B: Quits
MOR in other period 0.325*** -0.090***
(0.029) (0.023)
MOR in current period -0.103*** -0.278***
(0.023) (0.074)
Mean dep. var. 0.791 0.791 0.791
F-stat on excl instrument 124.6
Panel C: Fires
MOR in other period 0.325*** -0.061***
(0.029) (0.015)
MOR in current period -0.033** -0.188***
(0.014) (0.048)
Mean dep. var. 0.291 0.291 0.291
F-stat on excl instrument 124.6
Panel D: Regretted Quits
MOR in other period 0.325*** -0.075***
(0.029) (0.020)
MOR in current period -0.084*** -0.230***
(0.021) (0.065)
Mean dep. var. 0.621 0.621 0.621
F-stat on excl instrument 124.6
Panel E: Non-regretted Quits
MOR in other period 0.325*** -0.015
(0.029) (0.010)
MOR in current period -0.019** -0.048
(0.010) (0.030)
Mean dep. var. 0.169 0.169 0.169
F-stat on excl instrument 124.6

Notes: Standard errors clustered by manager in parentheses. An observation is an employee-month. In Panel

A, the dependent variable is whether an employee attrites in a given month. In the other panels, the
dependent variable is whether an employee experiences a particular type of attrition event in a given month.
All regressions include the same controls as in Table 2, plus current year dummies, the span of control for an
employee’s manager (plus a dummy for span being missing), and a 5th order polynomial in employee tenure.
Also, unlike Table 2, the controls are over time instead of for one month. * significant at 10%; ** significant
at 5%; *** significant at 1%

49
 Table 4: MOR and Employee Attrition: Exploiting New Joiners

Specification: 1st Stg OLS IV Reduced Form

Panel A: Attrition
MOR in other period 0.296*** -0.163
(0.043) (0.114)
MOR in current period -0.252** -0.550
(0.100) (0.370)
Mean dep. var. 1.446 1.446 1.446
F-stat on excl instrument 47.19
Panel B: Quits
MOR in other period 0.296*** -0.190**
(0.043) (0.093)
MOR in current period -0.212** -0.643**
(0.086) (0.308)
Mean dep. var. 0.864 0.864 0.864
F-stat on excl instrument 47.19
Panel C: Fires
MOR in other period 0.296*** -0.052
(0.043) (0.054)
MOR in current period -0.028 -0.175
(0.045) (0.180)
Mean dep. var. 0.362 0.362 0.362
F-stat on excl instrument 47.19
Panel D: Regretted Quits
MOR in other period 0.296*** -0.181**
(0.043) (0.088)
MOR in current period -0.190** -0.613**
(0.083) (0.292)
Mean dep. var. 0.798 0.798 0.798
F-stat on excl instrument 47.19
Panel E: Non-regretted Quits
MOR in other period 0.296*** -0.016
(0.043) (0.017)
MOR in current period -0.024* -0.053
(0.015) (0.057)
Mean dep. var. 0.0603 0.0603 0.0603
F-stat on excl instrument 47.19
Notes: This table is similar to Table 3, but restricts to new employees joining the firm after the
administration of the second survey (i.e., during period 2). * significant at 10%; ** significant at 5%; ***
significant at 1%

50
 Table 5: MOR and Employee Attrition: Exploiting New Joiners and People Switching
Managers

Specification: 1st Stg OLS IV Reduced Form

Panel A: Attrition
MOR in other period 0.280*** -0.093
(0.037) (0.069)
MOR in current period -0.157*** -0.332
(0.061) (0.236)
Mean dep. var. 1.512 1.512 1.512
F-stat on excl instrument 57.94
Panel B: Quits
MOR in other period 0.280*** -0.086
(0.037) (0.053)
MOR in current period -0.147*** -0.308*
(0.047) (0.185)
Mean dep. var. 0.880 0.880 0.880
F-stat on excl instrument 57.94
Panel C: Fires
MOR in other period 0.280*** -0.043
(0.037) (0.028)
MOR in current period -0.015 -0.153
(0.024) (0.097)
Mean dep. var. 0.327 0.327 0.327
F-stat on excl instrument 57.94
Panel D: Regretted Quits
MOR in other period 0.280*** -0.116**
(0.037) (0.047)
MOR in current period -0.127*** -0.412**
(0.045) (0.166)
Mean dep. var. 0.722 0.722 0.722
F-stat on excl instrument 57.94
Panel E: Non-regretted Quits
MOR in other period 0.280*** 0.029
(0.037) (0.019)
MOR in current period -0.020 0.105
(0.015) (0.073)
Mean dep. var. 0.158 0.158 0.158
F-stat on excl instrument 57.94
Notes: This table is similar to Table 3, but restricts to new employees joining the firm after the
administration of the second survey (i.e., during period 2) or to observations following a change in manager
during the second period (more precisely, to observations where a worker’s manager differs from the manager
they had during September Y1 when the first survey was administered). * significant at 10%; ** significant
at 5%; *** significant at 1%

51
 Table 6: Rothstein Test: Predicting Employe Outcomes Before Manager Switch as a Function of MOR of Future Manager

Dep. Var. Subjective Log Log Promoted Employee Log Key

performance salary salary x100 engagement stock individual
(normalized) x100 growth (normalized) grant (x100)
x100 holdings
x100
(1) (2) (3) (4) (5) (6) (7)
Panel A: OLS
MOR of future manager 0.037* -0.776** 0.294* 0.170 0.018* 2.326 0.785
measured in 2nd period (0.023) (0.346) (0.164) (0.139) (0.010) (1.553) (0.643)
Panel B: IV
MOR of future manager 0.121 -0.574 0.795 0.176 0.063* 8.495 -0.450
measured in 2nd period (0.088) (1.198) (0.781) (0.479) (0.037) (8.381) (1.636)
Panel C: Red. Form
52

MOR of future manager 0.034 -0.178 0.243 0.049 0.018* 1.571 -0.126
measured in 1st period (0.025) (0.373) (0.240) (0.135) (0.010) (1.512) (0.459)
Notes: Standard errors clustered by manager in parentheses. The controls are the same as in Table 3. The table presents regressions of employee
outcomes at the start of the sample as a function of the MOR of the employee’s new manager. The sample is restricted to employees who experience a
first change in manager during the second period. An observation is an employee-month occurring during period 1 (January Y1 -September Y1 ). Panel
A presents regressions of employee outcomes on the new manager’s MOR as measured during period 2. Panel B presents IV regressions of employee
outcomes on the new manager’s MOR as measured during period 2, while instrumenting using the new manager’s MOR as measured during period 1.
Panel C presents the reduced form regression of employee outcomes on the new manager’s MOR as measured during period 2. * significant at 10%; **
significant at 5%; *** significant at 1%
Table 7: MOR and Employee Attrition: Exploiting Managers Moving Across Locations and
Job Functions

Specification: OLS OLS IV IV

Panel A: Attrition
MOR of current manager in 1st period -0.305** -0.572**
(0.126) (0.250)
MOR of current manager in 2nd period -0.216** -0.680***
(0.105) (0.253)
Mean dep. var. 1.663 1.663 1.663 1.663
F-stat on excl instrument 42.01 31.10
Panel B: Quits
MOR of current manager in 1st period -0.108 -0.173
(0.080) (0.157)
MOR of current manager in 2nd period -0.065 -0.241
(0.067) (0.159)
Mean dep. var. 0.913 0.913 0.913 0.913
F-stat on excl instrument 42.01 31.10
Panel C: Fires
MOR of current manager in 1st period -0.115* -0.129
(0.060) (0.104)
MOR of current manager in 2nd period -0.049 -0.257**
(0.042) (0.118)
Mean dep. var. 0.223 0.223 0.223 0.223
F-stat on excl instrument 42.01 31.10
Panel D: Regretted Quits
MOR of current manager in 1st period -0.043 -0.192
(0.074) (0.145)
MOR of current manager in 2nd period -0.073 -0.096
(0.062) (0.145)
Mean dep. var. 0.727 0.727 0.727 0.727
F-stat on excl instrument 42.01 31.10
Panel E: Non-regretted Quits
MOR of current manager in 1st period -0.067* 0.016
(0.037) (0.072)
MOR of current manager in 2nd period 0.006 -0.150*
(0.030) (0.087)
Mean dep. var. 0.186 0.186 0.186 0.186
F-stat on excl instrument 42.01 31.10
Notes: This table presents regressions as in equation (6). An observation is a location-job function-period.
We use the raw locations with no groupings, and we exclude locations that have less than 10 worker-month
observations in the data provided before sample restrictions. The dependent variable is average attrition in
that cell. The regressor is the average MOR for managers in that cell, while measuring that manager’s MOR
in a particular period. All regressions include collapsed forms of the controls in Table 3. Instead of controlling
for current year, we control for period. * significant at 10%; ** significant at 5%; *** significant at 1%

53
 Table 8: MOR and Non-Attrition Outcomes

Dep. Var. Subjective Log Salary Promotion

performance Growth (x100) (x100)
(normalized)
(1) (2) (3) (4) (5) (6)
Panel A: OLS
MOR (normalized) 0.037*** -0.009 0.123 0.067 0.072 -0.048
(0.007) (0.008) (0.079) (0.102) (0.048) (0.116)
Panel B: IV
MOR (normalized) 0.095*** 0.032*** 0.064 0.157 -0.020 0.014
(0.023) (0.012) (0.205) (0.170) (0.135) (0.225)
Panel C: Red. Form
MOR (normalized) 0.031*** -0.015** 0.022 -0.086 -0.006 -0.007
(0.007) (0.007) (0.071) (0.115) (0.044) (0.107)
54

Employee FE No Yes No Yes No Yes

Notes: Standard errors clustered by manager in parentheses, with the exception of the even columns of Panel B, where standard errors are calculated
via block bootstrap (50 replications), with subsampling over managers. We do only 50 replications because the IV with fixed effects is
computationally demanding. The controls are the same as in Table 3. “Subjective performance” is an employee’s subjective performance on a 1-5
scale. In columns 5-6, the outcome is whether an employee receives a promotion, with coefficients multiplied by 100 for readability. * significant at
10%; ** significant at 5%; *** significant at 1%
 Table 9: What are Managers Rewarded For? Employees Survey Scores (MOR)

Dep var: Subjective Promoted Log Log Log Log Change in Key
performance (x100) salary salary stock change in span of individual
(normalized) (x100) growth grant stock control (x100)
(x100) holdings grants
(x100) (x100)
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A: OLS
MOR in current period 0.0641*** 0.101 -0.479 0.135 -2.685* 0.277 0.0637 -0.895
(0.0225) (0.0899) (0.403) (0.162) (1.543) (3.190) (0.0937) (0.863)
Panel B: IV
MOR in current period 0.419*** 0.673** -2.381 1.405** -0.670 5.874 0.260 2.078
(0.0870) (0.311) (1.505) (0.627) (5.252) (9.943) (0.303) (2.696)
Panel C: Red. Form
55

MOR in other period 0.138*** 0.221** -0.743* 0.462** -0.216 1.882 0.0843 0.682
(0.0222) (0.0950) (0.448) (0.193) (1.700) (3.173) (0.0986) (0.870)

Notes: Standard errors clustered by manager in parentheses. An observation is a manager-month. The controls are the same as in Table 3.
“Subjective performance” is a manager’s subjective performance on a 1-5 scale. “Promoted” is whether a manager receives a promotion in a given
month, with coefficients multiplied by 100 for readability. “Log salary growth” represents the change in a manager’s log salary from the present
month to one year ahead. “Stock grant holdings” measure the value of a person’s unvested sotck grants. “Log change in stock grants” uses the data
field from the firm on the value of new stock grants issued by the firm in the last year, and takes the log. “Change in span of control” represents the
change in a manager’s span of control from the present month to one year ahead. The “key individual” designation by the firm to individuals who are
deemed to especially important. * significant at 10%; ** significant at 5%; *** significant at 1%
“People Management Skills, Employee Attrition, and Man-
ager Rewards: Evidence from a High-Tech Firm”: Online
Appendix

Mitchell Hoffman Steven Tadelis

U. Toronto Rotman & NBER UC Berkeley Haas & NBER

The Online Appendix is organized as follows. Appendix A provides derivations and

additional results. For each subsection, we list the section of the main text that it accompanies.
Appendix B is our Data Appendix. Appendix C presents additional tables and figures.

1
A Additional Results
A.1 Comparison between Table 1 and Appendix Table C1 (Accom-
panying Section 2.4 from the Main Text)
Table 1 gives summary statistics for our main sample. In contrast, Appendix Table C1
provides summary statistics using the data before imposing that MOR be non-missing for a
worker’s manager in the current period and other period. We cannot give exact sample sizes
(to preserve firm confidentiality), but the sample in Appendix Table C1 is 2.2 times larger
than in Table 1, reflecting that MOR is missing for over half the managers in our dataset.
The largest differences between the two tables are that (1) the average span of control
is much lower in our main sample (5.10 vs. 9.35), reflecting that workers on small teams
are excluded from the sample because MOR is missing, and (2) workers in the main sample
experience fewer managers (reflecting that worker-months are excluded for certain managers).
There are some other smaller differences (e.g., a lower attrition rate in Table 1), but means on
several key variables are similar (i.e., subjective performance scores, MOR, share co-located
with managers).
To address potential concerns regarding selection bias, we repeat all of our main results
using two imputation strategies suggested by the firm, and obtain the same conclusions. These
imputation strategies are detailed below in Section B.

A.2 Econometric Derivations (Section 3.1)

OLS Derivation.

cov (yt , m
eτ)
plim(bbOLS ) =
var (m eτ)
2
βσm + βcov (m, uτ ) + cov (εt , m) + cov (εt , uτ )
= 2 + 2cov (m, u ) + σ 2
σm τ u
2
σm cov(εt , uτ ) cov(εt , m)
= 2 2
β+ 2 + 2
σm + σu σm + σu2 σm + σu2
σ2 cov(εt , uτ ) cov(εt , m)
plim(bbOLS − β) = − 2 u 2 β + 2 2
+ (9)
σ +σ σ +σ σ2 + σ2
| m{z u } | m {z u } | m {z u }
Attenuation Bias Contemp. Corr. ME Assignment Bias

where we used cov (m, uτ ) = 0 (Assumption 1) to go from the second line to the third line.

IV Derivation.

2
 cov (yt , m
e −t )
plim(bbIV ) =
cov (meτ, me −τ )
2
βσm + βcov (m, u−τ ) + cov (εt , m) + cov (εt , u−τ )
= 2 + cov (m, u ) + cov (m, u ) + cov (u , u )
σm τ −τ τ −τ
2
σm cov (εt , u−τ ) cov (εt , m)
= 2 + cov (u , u )
β + 2 + cov (u , u )
+ 2 + cov (u , u )
σm τ −τ σm τ −τ σm τ −τ
cov (uτ , u−τ ) cov (εt , u−τ ) cov (εt , m)
plim(bbIV − β) = − 2 β+ 2 + 2 (10)
σ + cov (uτ , u−τ ) σ + cov (uτ , u−τ ) σm + cov (uτ , u−τ )
| m {z } |m {z } | {z }
Attenuation Bias Asynchronously Corr. ME Assignment Bias

Reduced Form. For the reduced form expression in equation (5), the derivation is very
similar to those above for OLS and IV, so it is omitted for brevity.

A.3 What Happens When Manager Quality Varies Over Time?

(Section 3.2)
In Section 3, we present probability limits for OLS, IV, and reduced form estimators based on
the assumption that manager quality is fixed over time. Here, we derive the probability limits
of the estimators while allowing underlying manager quality to vary across the two periods in
our data. Suppose that yit = βmj,τ (t) + εit and write σ12 ≡ cov(mτ , m−τ ). We further assume
2
that var(m1 ) = var(m2 ) = σm . Under these assumptions, OLS is the same as when manager
quality is fixed. However, for IV, we have:

cov (yt , m
e −τ )
plim(bbIV ) =
cov (meτ, me −τ )
βcov(mτ , m−τ ) + βcov(mτ , u−τ ) + cov (εt , m−τ ) + cov (εt , u−τ )
=
cov(mτ , m−τ ) + cov(mτ , u−τ ) + cov(uτ , m−τ ) + cov (uτ , u−τ )
βσ12 cov(t , u−τ ) cov(t , m−τ )
= + +
σ12 + cov (uτ , u−τ ) σ12 + cov (uτ , u−τ ) σ12 + cov (uτ , u−τ )
cov (uτ , u−τ ) cov (εt , u−τ ) cov (εt , m)
plim(bbIV − β) = − β+ +
σ + cov (uτ , u−τ ) σ + cov (uτ , u−τ ) σ12 + cov (uτ , u−τ )
| 12 {z } | 12 {z } | {z }
Attenuation Bias Asynchronously Corr. ME Assignment Bias

2
Relative to the version with constant people management over time, the difference is that σm
is replaced by σ12 in the denominator. This makes attenuation bias worse. For the reduced

3
form, we have:

cov (yt , me −τ )
plim(bbRF ) =
var (me −τ )
2
σ12 − σm − σu2 cov(εt , u−τ ) cov(εt , m)
plim(bbRF − β) = 2 + σ2
β+ +
σm u σ2 + σ2 σ2 + σ2
| {z } | m {z u } | m {z u }
Attenuation Bias Asynchronously Corr. ME Assignment Bias

2
Attenuation bias is also worsened here the larger the divergence between σ12 and σm . However,
the formula is relatively similar.

A.4 Robustness Regarding Research Design Based on Workers Join-

ing the Firm or Changing Managers in the Second Period (Sec-
tion 4.3)
The analysis in Section 4.3 combines both workers joining the firm and worker changing
managers in the second period. As a robustness check, the analysis can also be performed
solely using incumbent workers changing managers in the second period. Appendix Table C10
shows that the results are broadly similar to those in Table 5, though with a few differences.
While the regretted quit coefficient is still statistically significantly negative, the non-regretted
quit coefficient is now significantly positive at the 10% level. Thus, while the overall quit
coefficient is still negative (as is the overall attrition coefficient), it is no longer statistically
significant.

A.5 Rothstein Test (Section 4.3)

The Rothstein test in Table 6 differs from our main analyses in that we are looking at an
employee’s early outcomes as a function of the people management skills of their future man-
agers. While the goal of the Rothstein test is to isolate the degree of assignment bias, there is
also a possibility that bias could arise due to attenuation bias or correlated measurement error.
Consider an OLS regression of early-measured employee outcomes on the MOR of a future
manager as measured during the second period. If a given employee is cheerful, there could
be bias if being cheerful makes the employee both more likely to achieve certain outcomes
in period 1, as well as more likely to rate his/her manager in a certain way. To overcome
this potential bias, as well as to address attenuation bias, we instrument the future manager’s
MOR as measured during the second period with the future manager’s MOR as measured
during the first period.
More concretely, consider an employee who changes from an initial manager (referred to
as the “old” manager) to a “new” manager. For OLS, we regress initial employee outcomes
on the MOR of the new manager during the second period. We obtain that:

4
 bbOLS = cov (me 1,new , yt )
var (m e 1,new )
cov (mnew + u1,new , βmold + εt )
=
var (mnew + u1 )
βcov (mnew , mold ) + cov (mnew , εt ) + βcov (u1,new , mold ) + cov (u1,new , εt )
= 2 + σ 2 + 2cov (u
σm u 1,new , mold )
1
= [βcov (mnew , mold ) + cov (mnew , εt ) + cov (u1,new , εt )]
σm + σu2
2

For IV, we have that:

bbIV cov (m
e 1,new , yt )
=
cov (m
e 1,new , m
e 2,new )
cov (mnew + u1,new , βmold + εt )
=
cov (mnew + u1 , mnew + u2 )
βcov (mnew , mold ) + cov (mnew , εt ) + βcov (u1,new , mold ) + cov (u1,new , εt )
= 2 + cov (u
σm 1,new , mold ) + cov (u2,new , mold ) + cov (u1 , u2 )
1
= 2 + cov (u , u )
[βcov (mnew , mold ) + cov (mnew , εt ) + cov (u1,new , εt )]
σm 1 2

Thus, instead of σu2 in the denominator, we have cov (u1 , u2 ) in the denominator, so we will
be less likely to suffer from attenuation bias. Provided that the two manager qualities are
uncorrelated over time (i.e., cov (mnew , mold ) = 0) and that cov (u1,new , εt ) = 0, then IV
should identify a coefficient which is proportional to cov (mnew , εt ), and is therefore a measure
of systematic assignment.

A.6 Managers Moving Across Locations or Job Functions (Section

4.4)
Locations and robustness. Locations are denoted in the data using a string variable. In
the dataset, there are cases of a large number of location string changes occurring during
the same month. Thus, there are some locations that only appear in period 1 and others
that only appear in period 2. These likely represent cases where either a location was simply
re-named in our dataset (without any physical movement of employees taking place) or where
an entire office re-located to another office building. To check that such instances do not drive
our results, Appendix Table C11 repeats the analysis in Table 3, but restricts to location-job
functions that occur during both periods in our data. For example, if there is a location that
suddenly seems to emerge in period 2 (perhaps due to a simple re-labeling of the building), that
“new location” will be removed from the sample in Appendix Table C11. The “old location”
will still be included for the months before the location string re-labeling occurred, and the
collapsed means (made from collapsing the employee-month panel) will take into account that
the old location was not observed for all of the second period.

5
 Comparison between our approach and Chetty et al. (2014). One difference
between our approach and Chetty et al. (2014) is that, because Chetty et al. (2014) have a
long panel, they can calculate leave-two-out VA estimators. For us, this is not feasible because
we only have two waves of the survey.1

A.7 Assessing Coefficient Stability when Adding Richer Controls

using Oster Test
To assess coefficient stability, we consider the test of Oster (2017), who builds on Altonji et
al. (2005). Oster (2017) presents her test using OLS regressions. To adopt the test to our IV
setting, we follow Enikolopov et al. (2017) and perform the Oster test using the reduced form.
The idea of the Oster (2017) test is to compare the degree of coefficients movements
with the amount of movement in R-squared values. We take the IV regressions reported in
Tables C12-C14 and Table C23, and perform the reduced form regressions instead. Column 1
represents the specification with base controls, whereas column 5 represent the specification
with full controls. Following Oster (2017), we assume a maximum R-squared value that is 1.3
times the R-squared with the fullest controls (i.e., the column 5 specifications for us).
Following Oster (2017), we calculate values of δ, which represent the ratio of selection
on unobservables relative to selection observables that would be required in order the true
coefficient to not be in the observed direction. Oster (2017) argues that estimated δ values of
one or greater provide evidence of coefficient stability. In addition, δ coefficients less than 0
suggest that the true, bias-adjusted coefficients are larger than the estimated ones (Satyanath
et al., 2017). In all cases, we obtain δ values either greater than 1 or less than 0, thereby
strengthening our evidence for robustness.

A.8 Heterogeneity Analysis (Section 4.5)

Geography. There has been a lot of recent interest in how management varies across coun-
tries, particularly in rich vs. poor countries. Bloom et al. (2014) document that management
practices are substantially better in richer countries than in poorer countries. Hoffman et
al. (2018) document that frontline supervisors appear to matter more in rich countries than
in poor countries for the case of employee attrition. Most of the workers we study are in
rich countries (particularly the US), but there are roughly 10% of employee records in “poor
countries” (China, India, and Malaysia). In our setting, in Table C16, MOR has a stronger
negative relation for US workers than for non-US workers, and particularly so compared to
non-US poor countries (though the instrument is weak with F = 3 for workers in poor, foreign
countries at the firm). However, the presence of relatively large standard errors implies that
our comparisons are only suggestive.
Occupation. Appendix Table C17 examines heterogeneity in attrition results by oc-
cupation / job function. IV estimates are relatively similar across workers in engineering,
marketing, and finance.
1
Recall from Section 2.3 of the main text that we were provided data from a third survey wave, but it is
missing for a large business unit at the firm and has some different survey questions from the other two waves.

6
 Hierarchy. Appendix Table C18 shows that the negative relation between MOR and
attrition actually appears to be larger for individuals toward the upper part of the firm hier-
archy. We divide individuals at the firm into three levels of hierarchy according to their salary
grade, following how the firm often segments employees in its internal reporting. It is natural
to analyze heterogeneity in manager effects by hierarchy, as theories of managers emphasize
different roles for managers at different levels of hierarchy. For example, in knowledge-based
theories of the firm (Garicano, 2000), managers solve increasingly complex problems as they
ascend the firm hierarchy. However, as for occupation and geography, the relatively large
standard errors limit our ability to draw sharp comparisons in IV estimates by hierarchy.

A.9 Further Discussion on VA (Section 4.5)

Section 4.5 of the main text discusses our split sample approach for estimating the variance
of manager fixed effects in retention. (Recall that fixed effects in retention constitute our
measure of manager value-added (VA).) Assuming that the sampling error is uncorrelated
across samples and with underlying VA, the covariance of the estimated manager fixed effects
across the two samples is equal to the variance of underlying VA.
Beyond dealing with sampling error, this approach may also be useful in addressing
correlated shocks. For example, if a manager’s team is assigned to do a good project, this
might cause the manager to appear to matter a lot for reducing attrition, whereas it is the
good project that is reducing attrition. Thus, an advantage of splitting across the two periods
(instead of randomly splitting the sample) is that doing so might do a better job of addressing
transient correlated shocks. However, a disadvantage of splitting across the two periods is
that the two periods are uneven in size; the second period lasts 18 months, compared to 9
months for the first period.

A.10 Rewards Results, Including Manager VA as a Regressor (Sec-

tion 6.2)
As seen in Table C25, we see that the relation of MOR to rewards is mostly similar when
controlling for manager VA. An additional difference (relative to Table 9) is that the promotion
results are no longer statistically significant, but the magnitudes are very similar.
Beyond the exercise in Table C25, we can additionally include a manager’s subjective
performance as a regressor, with the goal of seeing whether “managing up” is rewarded more
at the firm than “managing down.” Appendix Table C26 shows the results of these regressions.
While adding manager’s subjective performance shrinks the coefficients on MOR, we do not
emphasize these regressions for two reasons. First, subjective performance may be a “bad
control” (Angrist and Pischke, 2008), as it a critical means by which MOR affects the other
variables. Second, subjective performance scores can be given taking into account a manager’s
MOR, so it is not surprising that subjective performance scores are more predictive.

7
B Data Appendix
Data assembly. We were provided two main datasets for our project. First, we received
the main employee-month personnel dataset that was assembled for us by an analyst at the
high-tech firm. A variety of files were combined together during this process. The analyst
also subjected the data to cleaning. Second, we received manager-level results of the three
employee surveys.
Manager survey variables. As described in Section 2.3 of the main text, manager
survey scores are not collected when workers are part of small teams.
One concern is that selection bias from missing MOR data could affect our results. To
address this concern, we pursued two different imputation strategies, and our conclusions were
unaffected by both.
In the first strategy, we filled in missing MOR scores using “roll-up survey values” (when
available). Roll-up scores are manager scores using all the individuals under a given manager
in the organization. A rationale for using roll-up scores in imputing MOR is that excellent
people management skills may flow downward in an organization, making one’s employees
better people managers. We pursued the strategy of using roll-up scores, as this is what our
study firm often does for reporting purposes. In the second strategy, we take advantage of
the fact that the cleaned employee-month personnel dataset we were provided also contains
MOR scores that have been subject to various analyst-added imputations, including the roll-
up survey imputations. After starting with these MOR variables, we filled in missing values
using the raw scores, and if still missing, using the roll-up scores.
As a robustness check, we performed all our analyses using both strategies. Doing so
leads to results that are very similar and that are even sharper/more precise than our reported
results, reflecting a larger sample size. This suggests that missing data on MOR does not drive
our conclusions.
Our analysis is done using manager overall rating (“MOR”). This is calculated by
normalizing MOR separately by period. We note also that MOR is an acronym created by
the authors—the firm usually refers to the score as the manager effectiveness score.
Regretted and non-regretted attrition. A manager from HR told us that the data
field in our data about whether a quit was regretted or non-regretted may not have always
been recorded in the same manner, and may have changed over time. Usually, the data would
be entered by a person’s former manager. However, it can also be that the data field would
incorporate information from an HR business partner who conducted an exit interview of the
former employee. Furthermore, the manager informed us that the regretted/non-regretted
field could also sometimes be “algorithmic” based on the subjective performance scores of the
former employee.2
Thus, some caution is warranted in interpreting our results on regretted and non-
regretted attrition. Still, whether the classification is done by a manager or using subjec-
tive performance data, it still reflects a desire to classify attrition as good or bad from the
perspective of the firm.
2
The HR manager also did not see certain whether it was that the method of scoring had changed over
time or whether the computer default had changed over time. Throughout the data, regretted quits are more
common than non-regretted quits.

8
 Our time fixed effects adjust for possible changes over time in how regretted/non-
regretted was classified.3 Importantly, however, our conversations with the firm gave us no
reason to be concerned that whether a quit was classified as regretted or non-regretted would
be mechanically related to or correlated with whether a manager had good people management
skills.
Salary. While workers at the firm are paid in different currencies, we restrict our salary
analyses in the paper to workers paid in US dollars. However, we checked that our results
on MOR and salary increases in Table 8 are robust to including all workers. To do this, we
convert salaries to US dollars using the exchange rate as of March 1 of year Y2 (which falls in
the middle of our data period).
Stock grants and holding power. Only individuals sufficiently high up in the cor-
porate hierarchy (i.e., with sufficiently high salary grades) are eligible to receive stock grants.
The required level varies between technical and non-technical jobs. In the data that was
provided to us, the holding power variable is missing for a lot of observations, presumably
reflecting that many employees are not eligible to receive stock grants. Our analysis of stock
grants is performed only using observations that are non-missing. (That is, we do not assign
zero values to observations where holding power is missing.)
Key individual. Persons at the firm who are recognized as an integral part of the
company are designated “key individuals.” The firm uses a slightly different term to refer to
such persons, but we have modified it for the paper to preserve firm confidentiality.

B.1 Summary of Sample Restrictions

1. In cleaning our employee-month panel, we exclude observations sharing the same person
ID and month (dropping 1% of observations).

2. To focus on high-skill workers, we eliminate worker records in the job function of cus-
tomer service / operations (dropping 32% of observations relative to the start).

3. We exclude observations occurring in April and May of Y3 , as the location identifiers

change during this period (dropping 4% of observations relative to the start).

4. We exclude workers for whom the manager does not have MOR in both the current and
the other period (dropping 34% of observations relative to the start). We require that
both MOR in the current period and MOR in the other period be observed in order to
perform our main IV analysis.

B.2 Y3 Survey Questions

The survey questions were slightly different for the Y3 . They are listed below.

1. My immediate manager provides ongoing coaching and guidance on how I can improve
my performance.
3
In a regression of whether a quit was regretted on year dummies and other basic controls, the share of
quits which are regretted is 3pp higher in Y2 than Y1 and is 9pp higher in Y3 than Y1 .

9
 2. My immediate manager actively supports my efforts regarding professional / career
development.

3. My immediate manager extends influence and leadership across organizational bound-

aries.

4. My immediate manager creates the conditions that support stronger engagement at

work.

5. I would recommend my manager to others.

C Additional Figures and Tables

10
 Table C1: Summary Statistics for Dataset before Imposing Restriction of Non-missing
MOR for Managers in the Current and Other Period

Panel A: Overall numbers

Share of records, employee in US 0.66
Share of records from managers 0.20
Share of records for engineers 0.33
Co-located with manager 0.81
Same function as manager 0.85
Average manager span (employees/mgr) 5.10
Managers per employee 2.27
Managers per employee (weighted by tenure) 2.56

Panel B: Several outcomes and

regressors of interest
Variable: mean sd min max
Attrition probability (monthly) x100 1.55 12.36 0 100
Quit probability (monthly) x100 0.87 9.28 0 100
Fire probability (monthly) x100 0.34 5.82 0 100
Regretted quit prob (monthly) x100 0.69 8.27 0 100
Non-regretted quit prob (monthly) x100 0.18 4.22 0 100
Subjective performance rating 3.32 .81 1 5
Log salary Confidential
Promotion probability (monthly) Confidential
Manager overall rating 80.25 15.87 0 100
Manager gives clear expectations 82.75 17.44 0 100
Manager provides coaching 73.94 21.34 0 100
Manager supports career dev 76.54 20.02 0 100
Manager involves people 83.77 17.16 0 100
Manager instills poz attitude 82.15 19.2 0 100
Manager is someone I trust 82.01 18.22 0 100
Notes: This table is similar to Table 1. The difference is that it summarizes the data before imposing the
restriction of a worker’s manager having non-missing MOR in both the current and other period.

11
 Figure C1: Correlation of Survey Items across the Two Waves

85
90

80
85

75
expec in P2

coach in P2
80

70
75

65
Coefficient = 0.263 Coefficient = 0.306
(0.031) (0.028)
70

60
40 60 80 100 20 40 60 80 100
expec in P1 coach in P1

(a) Manager gives clear expectations (b) Manager provides coaching

85

90
80

85
crdev in P2

invol in P2
75

80
70

75
65

Coefficient = 0.302 Coefficient = 0.302

(0.027) (0.034)
70

20 40 60 80 100 40 60 80 100
crdev in P1 invol in P1

(c) Manager supports career development (d) Manager involves people

90

90
85

85
80
pozat in P2

trstw in P2
80
75

75
70

70
65

Coefficient = 0.410 Coefficient = 0.334

(0.035) (0.031)
40 60 80 100 40 60 80 100
pozat in P1 trstw in P1

(e) Manager instills positive attitude (f) Manager is someone I trust

Notes: These graphs are similar to Figure 1 in the main text. The difference is that these are graphs for the
six individual manager questions (as opposed to MOR).

12
 Figure C2: Reduced Form Binned Scatter Plots: Exploiting New Joiners
2.5

1.5
2

1
Coef

Coef
1.5

.5
1

Coefficient = -0.163 Coefficient = -0.190

(0.114) (0.093)
.5

0
-2 -1 0 1 2 -2 -1 0 1 2
MOR in Other Period MOR in Other Period

(a) Attrition (b) Quits

1.5
.8
.6

1
Coef

Coef
.4

.5
.2

Coefficient = -0.052 Coefficient = -0.181

(0.054) (0.088)
0

-2 -1 0 1 2 -2 -1 0 1 2
MOR in Other Period MOR in Other Period

(c) Fires (d) Regretted Quits

14
.2
.15

12
.1
Coef

Coef
10
.05

8
0

Coefficient = -0.016 Coefficient = -0.783

-.05

(0.017) (0.509)
6

-2 -1 0 1 2 -2 -1 0 1 2
MOR in Other Period MOR in Other Period

(e) Non-regretted Quits (f) Worker changes manager

Notes: This figure is similar to Figure 2 in the main text. The difference is that these figures are made for
the joiners analysis. That is, the regressions correspond to those in Table 4.

13
 Figure C3: Reduced Form Binned Scatter Plots: Exploiting New Joiners and People
Switching Managers

1.4
2
1.8

1.2
1.6
Coef

Coef
1
1.4

.8
1.2

Coefficient = -0.093 Coefficient = -0.086

.6
1

(0.069) (0.053)
-2 -1 0 1 2 -2 -1 0 1 2
MOR in Other Period MOR in Other Period

(a) Attrition (b) Quits

1.2
.5

1
.4
Coef

Coef
.8
.3

.6
.2

Coefficient = -0.043 Coefficient = -0.116

(0.028) (0.047)
.1

.4

-2 -1 0 1 2 -2 -1 0 1 2
MOR in Other Period MOR in Other Period

(c) Fires (d) Regretted Quits

12
.3
.25

11
10
.2
Coef

Coef
.15

9
.1

Coefficient = 0.029 Coefficient = -0.509

(0.019) (0.399)
.05

-2 -1 0 1 2 -2 -1 0 1 2
MOR in Other Period MOR in Other Period

(e) Non-regretted Quits (f) Worker changes manager

Notes: This figure is similar to Figure 2 in the main text. The difference is that these figures are made for
the joiners analysis. That is, the regressions correspond to those in Table 5.

14
 Table C2: Manager Characteristics, Correlation Table

Variables: Clear Coaching Career dev Involves Positive Someone

expectations people attitude I trust
Manager gives clear expectations 1
Manager provides coaching 0.664 1
Manager supports career development 0.580 0.706 1
Manager involves people 0.573 0.554 0.601 1
Manager instills positive attitude 0.575 0.569 0.589 0.675 1
Manager is someone I trust 0.634 0.587 0.633 0.676 0.728 1
Notes: Correlation coefficients are reported. The table uses data from employee responses to the 6 questions about managers in Section 2.3. An
observation is a manager-survey (i.e., with two survey periods per manager).
15
 Table C3: Principal Component Analysis

Variables: Component Component Component Component

1 2 3 4
Eigenvalue 4.33 0.58 0.42 0.34
Proportion variance explained 0.69 0.10 0.07 0.06
Manager gives clear expectations 0.40 0.29 0.77 0.15
Manager provides coaching 0.40 0.56 -0.09 -0.05
Manager supports career development 0.41 0.35 -0.60 -0.04
Manager involves people 0.40 -0.41 -0.15 0.79
Manager instills positive attitude 0.41 -0.45 -0.01 -0.49
Manager is someone I trust 0.42 -0.32 0.09 -0.34
Notes: This table presents the results of the principal components analysis. The table uses data from employee responses to the 6 questions about
managers in Section 2.3. An observation is a manager-survey (i.e., with two survey periods per manager).
16
 Table C4: Transition Matrix, MOR, by Quintile

Q1 Q2 Q3 Q4 Q5
in Y2 in Y2 in Y2 in Y2 in Y2
1st Quintile in Y1 .37 .28 .18 .1 .07
2nd Quintile in Y1 .27 .27 .19 .17 .1
3rd Quintile in Y1 .2 .24 .2 .22 .13
4th Quintile in Y1 .1 .15 .23 .25 .28
5th Quintile in Y1 .09 .09 .21 .21 .39
Notes: This table uses the data from the 6 questions from employees about their managers in Section 2.3. The numbers represent the share of
managers in a given MOR quintile during Y1 who transition to a particular MOR quintile during Y2 . Higher quintiles represent higher underlying
MOR scores. For example, the number in the first row and fourth column indicates, among managers in the 1st MOR quintile (i.e., lowest 20% of
scores) during Y1 , the share that achieve a score in the 4th MOR quintile (i.e., the second highest quintile) during Y2 .
17
 Table C5: Robustness Analysing on Persistence of Managerial Characteristics: Using all the Manager Characteristics as
Regressors at the Same Time

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

Dep. Variables: Clear Coaching Career dev Involves Positive Someone
expectations people attitude I trust
Manager sets clear expectations 0.15*** 0.03 -0.02 0.08 -0.02 -0.01
(0.05) (0.05) (0.05) (0.05) (0.05) (0.05)
Manager gives coaching 0.09* 0.23*** 0.14*** 0.00 -0.01 0.01
(0.06) (0.05) (0.05) (0.05) (0.06) (0.05)
Manager promotes career development 0.06 0.09* 0.19*** 0.04 0.06 0.08
(0.05) (0.05) (0.05) (0.06) (0.06) (0.05)
Manager involves people 0.01 0.00 -0.00 0.12** 0.02 0.04
(0.05) (0.05) (0.05) (0.05) (0.05) (0.05)
Manager instills a positive attitude 0.00 -0.05 0.06 0.13** 0.36*** 0.13**
(0.05) (0.05) (0.05) (0.06) (0.06) (0.06)
18

Employees trust the manager 0.01 0.03 0.02 0.02 0.06 0.18***
(0.06) (0.06) (0.06) (0.05) (0.06) (0.06)
R-squared 0.237 0.255 0.266 0.230 0.277 0.252
Notes: This table is a robustness check to Table 2. Instead of regressing a particular Y2 characteristic on the same characteristic in Y1 and various
controls, we regress each Y2 characteristics on all the Y1 characteristics at once (plus controls). * significant at 10%; ** significant at 5%; ***
significant at 1%
 Table C6: Robustness Analysing on Persistence of Managerial Characteristics: Using the
Y3 Survey

(1) (2)
Variables: Overall Overall
MOR MOR
Sample: Y3 Y1 , Y2 , Y3
Lagged MOR 0.22*** 0.27***
(0.08) (0.05)
Notes: This table is a robustness check to Table 2. The difference is that we use all three surveys (in
Y1 , Y2 , Y3 ) as opposed to just the Y1 and Y2 surveys. The sample is restricted to managers for whom we
observe all three surveys. Column 1 analyzes MOR in Y3 as a function of MOR in Y2 . Column 2 analyzes
MOR in Y2 and Y3 as a function of the MOR in the previous period. * significant at 10%; ** significant at
5%; *** significant at 1%

19
 Table C7: Summary of Identification Strategies

Strategy: Rationales
Baseline IV -Eliminates attenuation bias (if survey measurement error is uncorrelated across periods).
-Replaces contemporaneously correlated measurement error with asynchronously correlated measurement error.

Joiners -Setting with little bias from asynchronously correlated measurement error. Likely zero bias if common
component of survey measurement error (u−τ ) and quitting equation error term (εt ) is not persistent.
-Setting where assignment bias is likely very small.

Workers switching -Broader sample than joiners (good for external validity).
managers or joiners -Can formally test for assignment bias.
-Can examine time path of people management effects (separate from tenure effects).
-Less risk of asynchronously correlated measurement error than baseline IV.

Managers switching -Eliminates concern about persistent, common component of survey measurement
20

locations or functions error (u−τ ) and quitting equation error term (εt ).
-Setting where assignment bias is likely very small.
Notes: This table lists the different identification strategies in the paper, as well as rationales for them.
 Table C8: MOR and Employee Attrition: High- vs. Low-Productivity Employees

Specification: 1st Stg OLS IV Reduced Form

Panel A: All attrition, high-productivity employees
MOR in other period 0.318*** -0.070**
(0.030) (0.029)
MOR in current period -0.099*** -0.221**
(0.031) (0.093)
Mean dep. var. 0.655 0.655 0.655
F-stat on excl instrument 110.0
Panel B: All attrition, low-productivity employees
MOR in other period 0.326*** -0.185***
(0.029) (0.046)
MOR in current period -0.118*** -0.566***
(0.044) (0.148)
Mean dep. var. 1.540 1.540 1.540
F-stat on excl instrument 129.1

Notes: The panels in this table are similar to Panel A of Table 3. The difference is that we split the sample
based on whether employees are “high” or “low” productivity individuals. Workers are classified as high or
low productivity based on subjective performance scores, as described in Section 4.1. * significant at 10%; **
significant at 5%; *** significant at 1%

21
 Table C9: MOR and Whether an Employee Gets Changed to a New Manager

Specification: 1st Stg OLS IV Reduced Form

MOR in other period 0.326*** -0.443**
(0.029) (0.174)
MOR in current period -0.767*** -1.361**
(0.182) (0.547)
Mean dep. var. 6.087 6.087 6.087
F-stat on excl instrument 125.3

Notes: This table is similar to Panel A of Table 3. The difference is that instead of analyzing attrition, we
analyze whether an employee changes to a different manager in the next month (with coefficients multiplied
by 100 for ease of exposition). For example, the table examines whether the MOR of an employee’s manager
in January Y1 predicts whether January is the last month that the employee is supervised by that manager
(i.e., the manager ID for February Y1 is different from that in January Y1 ). * significant at 10%; **
significant at 5%; *** significant at 1%

22
Table C10: Robustness Check on Exploiting New Joiners and People Switching Managers:
Only Analyze People Switching Managers

Specification: 1st Stg OLS IV Reduced Form

Panel A: Attrition
MOR in other period 0.267*** -0.061
(0.040) (0.072)
MOR in current period -0.135* -0.228
(0.071) (0.265)
Mean dep. var. 1.543 1.543 1.543
F-stat on excl instrument 44.50
Panel B: Quits
MOR in other period 0.267*** -0.040
(0.040) (0.054)
MOR in current period -0.137*** -0.150
(0.053) (0.202)
Mean dep. var. 0.890 0.890 0.890
F-stat on excl instrument 44.50
Panel C: Fires
MOR in other period 0.267*** -0.034
(0.040) (0.030)
MOR in current period -0.004 -0.127
(0.025) (0.111)
Mean dep. var. 0.311 0.311 0.311
F-stat on excl instrument 44.50
Panel D: Regretted Quits
MOR in other period 0.267*** -0.092*
(0.040) (0.049)
MOR in current period -0.123** -0.345*
(0.049) (0.187)
Mean dep. var. 0.688 0.688 0.688
F-stat on excl instrument 44.50
Panel E: Non-regretted Quits
MOR in other period 0.267*** 0.052*
(0.040) (0.027)
MOR in current period -0.014 0.195*
(0.020) (0.107)
Mean dep. var. 0.202 0.202 0.202
F-stat on excl instrument 44.50
Notes: This table is similar to Table 5, but restricts only to observations following a change in manager
during the second period (more precisely, to observations where a worker’s manager differs from the manager
they had during September Y1 when first survey was administered). That is, we exclude workers who join
the firm during the second period.* significant at 10%; ** significant at 5%; *** significant at 1%

23
 Table C11: Robustness Check on Exploiting Managers Moving Across Locations and Job
Functions: Restrict to Location-Job Functions in Both Periods of the Data

Specification: OLS OLS IV IV

Panel A: Attrition
MOR of current manager in 1st period -0.267** -0.738**
(0.135) (0.290)
MOR of current manager in 2nd period -0.271** -0.622**
(0.116) (0.283)
Mean dep. var. 1.458 1.458 1.458 1.458
F-stat on excl instrument 29.29 22.33
Panel B: Quits
MOR of current manager in 1st period -0.076 -0.237
(0.083) (0.178)
MOR of current manager in 2nd period -0.087 -0.178
(0.074) (0.174)
Mean dep. var. 0.722 0.722 0.722 0.722
F-stat on excl instrument 29.29 22.33
Panel C: Fires
MOR of current manager in 1st period -0.123* -0.195*
(0.065) (0.116)
MOR of current manager in 2nd period -0.072 -0.286**
(0.045) (0.132)
Mean dep. var. 0.238 0.238 0.238 0.238
F-stat on excl instrument 29.29 22.33
Panel D: Regretted Quits
MOR of current manager in 1st period -0.029 -0.244
(0.078) (0.166)
MOR of current manager in 2nd period -0.090 -0.067
(0.069) (0.162)
Mean dep. var. 0.606 0.606 0.606 0.606
F-stat on excl instrument 29.29 22.33
Panel E: Non-regretted Quits
MOR of current manager in 1st period -0.050 -0.001
(0.040) (0.076)
MOR of current manager in 2nd period -0.000 -0.116
(0.031) (0.097)
Mean dep. var. 0.116 0.116 0.116 0.116
F-stat on excl instrument 29.29 22.33
Notes: This table is a robustness check to Table 7. The difference is that we restrict attention to location-job
functions that are observed during both periods. * significant at 10%; ** significant at 5%; *** significant at
1%

24
 Table C12: Robustness for IV in Table 3: Gradually Adding Additional Controls

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

Panel A: Attrition
MOR in current period -0.475*** -0.481*** -0.487*** -0.499*** -0.495***
(0.103) (0.104) (0.105) (0.106) (0.107)
Panel B: Quits
MOR in current period -0.278*** -0.283*** -0.281*** -0.285*** -0.281***
(0.0741) (0.0750) (0.0752) (0.0754) (0.0754)
Panel C: Fires
MOR in current period -0.188*** -0.190*** -0.193*** -0.200*** -0.201***
(0.0483) (0.0490) (0.0498) (0.0501) (0.0501)
25

Panel D: Regretted Quits

MOR in current period -0.230*** -0.235*** -0.234*** -0.236*** -0.234***
(0.0648) (0.0656) (0.0661) (0.0664) (0.0664)
Panel E: Non-regretted Quits
MOR in current period -0.0475 -0.0485 -0.0465 -0.0485 -0.0471
(0.0298) (0.0301) (0.0299) (0.0303) (0.0303)
Base Controls Y Y Y Y Y
Business Unit X Job Function Dummies N Y Y Y Y
Business Unit X Salary Grade Dummies N N Y Y Y
Job Function X Salary Grade Dummies N N N Y Y
Current Month Dummies N N N N Y
Notes: Standard errors clustered by manager in parentheses. This table is a robustness check to Table 3. It takes the IV specifications in the 5 panels
and gradually adds additional control variables. * significant at 10%; ** significant at 5%; *** significant at 1%
 Table C13: Robustness for IV in Table 4: Gradually Adding Additional Controls

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

Panel A: Attrition
MOR in current period -0.550 -0.533 -0.535 -0.573 -0.576
(0.370) (0.376) (0.384) (0.408) (0.408)
Panel B: Quits
MOR in current period -0.643** -0.636** -0.617* -0.677** -0.670**
(0.308) (0.312) (0.317) (0.341) (0.341)
Panel C: Fires
MOR in current period -0.175 -0.166 -0.159 -0.156 -0.173
(0.180) (0.179) (0.179) (0.192) (0.193)
26

Panel D: Regretted Quits

MOR in current period -0.613** -0.607** -0.600** -0.671** -0.664**
(0.292) (0.296) (0.302) (0.326) (0.325)
Panel E: Non-regretted Quits
MOR in current period -0.0526 -0.0511 -0.0406 -0.0334 -0.0336
(0.0567) (0.0534) (0.0478) (0.0469) (0.0473)
Base Controls Y Y Y Y Y
Business Unit X Job Function Dummies N Y Y Y Y
Business Unit X Salary Grade Dummies N N Y Y Y
Job Function X Salary Grade Dummies N N N Y Y
Current Month Dummies N N N N Y
Notes: Standard errors clustered by manager in parentheses. This table is a robustness check to Table 4. It takes the IV specifications in the 5 panels
and gradually adds additional control variables. * significant at 10%; ** significant at 5%; *** significant at 1%
 Table C14: Robustness for IV in Table 5: Gradually Adding Additional Controls

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

Panel A: Attrition
MOR in current period -0.332 -0.317 -0.314 -0.336 -0.340
(0.236) (0.235) (0.235) (0.243) (0.244)
Panel B: Quits
MOR in current period -0.308* -0.289 -0.295 -0.245 -0.248
(0.185) (0.182) (0.181) (0.186) (0.187)
Panel C: Fires
MOR in current period -0.153 -0.148 -0.144 -0.194** -0.198**
(0.0971) (0.0956) (0.0968) (0.0958) (0.0964)
27

Panel D: Regretted Quits

MOR in current period -0.412** -0.393** -0.406** -0.365** -0.368**
(0.166) (0.162) (0.161) (0.165) (0.166)
Panel E: Non-regretted Quits
MOR in current period 0.105 0.104 0.112 0.120* 0.120*
(0.0728) (0.0721) (0.0707) (0.0721) (0.0721)
Base Controls Y Y Y Y Y
Business Unit X Job Function Dummies N Y Y Y Y
Business Unit X Salary Grade Dummies N N Y Y Y
Job Function X Salary Grade Dummies N N N Y Y
Current Month Dummies N N N N Y
Notes: Standard errors clustered by manager in parentheses. This table is a robustness check to Table 5. It takes the IV specifications in the 5 panels
and gradually adds additional control variables. * significant at 10%; ** significant at 5%; *** significant at 1%
 Table C15: Robustness for IV in Table 7: Gradually Adding Additional Controls

(1) (2) (3) (4) (5) (6) (7) (8)

Panel A: Attrition
MOR of current manager in 1st period -0.572** -0.782*** -0.759*** -0.645***
(0.250) (0.262) (0.246) (0.239)
MOR of current manager in 2nd period -0.680*** -0.727*** -0.593** -0.581**
(0.253) (0.261) (0.249) (0.256)
Panel B: Quits
MOR of current manager in 1st period -0.173 -0.278 -0.239 -0.220
(0.157) (0.175) (0.181) (0.184)
MOR of current manager in 2nd period -0.241 -0.280* -0.255 -0.260
(0.159) (0.165) (0.160) (0.168)
Panel C: Fires
MOR of current manager in 1st period -0.129 -0.149 -0.122 -0.0604
(0.104) (0.112) (0.122) (0.121)
28

MOR of current manager in 2nd period -0.257** -0.291** -0.235* -0.213*

(0.118) (0.135) (0.130) (0.127)
Panel D: Regretted Quits
MOR of current manager in 1st period -0.192 -0.274* -0.266 -0.250
(0.145) (0.160) (0.167) (0.166)
MOR of current manager in 2nd period -0.0961 -0.124 -0.122 -0.132
(0.145) (0.152) (0.149) (0.150)
Panel E: Non-regretted Quits
MOR of current manager in 1st period 0.0158 -0.00766 0.0250 0.0311
(0.0724) (0.0747) (0.0776) (0.0867)
MOR of current manager in 2nd period -0.150* -0.159* -0.136 -0.130
(0.0870) (0.0929) (0.0926) (0.0931)
Base Controls Y Y Y Y Y Y Y Y
Business Unit X Job Function Dummies N Y Y Y N Y Y Y
Business Unit X Salary Grade Dummies N N Y Y N N Y Y
Job Function X Salary Grade Dummies N N N Y N N N Y

Notes: Standard errors clustered by manager in parentheses. This table is a robustness check to Table 7. It takes the two IV specifications in the 5
panels and gradually adds additional control variables. * significant at 10%; ** significant at 5%; *** significant at 1%
 Table C16: MOR and Employee Attrition: Heterogeneity by Geography

Specification: 1st Stg OLS IV Reduced Form

Panel A: Domestic
MOR in other period 0.348*** -0.205***
(0.033) (0.041)
MOR in current period -0.189*** -0.588***
(0.039) (0.128)
Mean dep. var. 1.435 1.435 1.435
F-stat on excl instrument 109.5
Panel B: Foreign
MOR in other period 0.263*** -0.068
(0.052) (0.047)
MOR in current period -0.103** -0.258
(0.051) (0.181)
Mean dep. var. 1.231 1.231 1.231
F-stat on excl instrument 25.90
Panel C: Foreign, Poor Country
MOR in other period 0.137* 0.004
(0.079) (0.077)
MOR in current period -0.097 0.029
(0.095) (0.564)
Mean dep. var. 1.286 1.286 1.286
F-stat on excl instrument 2.974
Panel D: Foreign, Rich Country
MOR in other period 0.311*** -0.096*
(0.067) (0.056)
MOR in current period -0.096* -0.308
(0.058) (0.188)
Mean dep. var. 1.197 1.197 1.197
F-stat on excl instrument 21.88

Notes: Each panel is similar to Panel A of Table 3. The difference is that we examine heterogeneity by
geography. * significant at 10%; ** significant at 5%; *** significant at 1%

29
 Table C17: MOR and Employee Attrition: Heterogeneity by Occupation

Specification: 1st Stg OLS IV Reduced Form

Panel A: Engineers
MOR in other period 0.190*** -0.055
(0.043) (0.046)
MOR in current period -0.163*** -0.288
(0.045) (0.244)
Mean dep. var. 1.231 1.231 1.231
F-stat on excl instrument 20.08
Panel B: Marketing
MOR in other period 0.343*** -0.170**
(0.084) (0.080)
MOR in current period -0.012 -0.495*
(0.068) (0.269)
Mean dep. var. 1.284 1.284 1.284
F-stat on excl instrument 16.66
Panel C: Finance
MOR in other period 0.309*** -0.204*
(0.087) (0.114)
MOR in current period -0.405*** -0.660*
(0.137) (0.383)
Mean dep. var. 1.094 1.094 1.094
F-stat on excl instrument 12.51

Notes: Each panel is similar to Panel A of Table 3. The difference is that we examine heterogeneity by
worker occupation (job function). * significant at 10%; ** significant at 5%; *** significant at 1%

30
 Table C18: MOR and Employee Attrition: Heterogeneity by Hierarchy

Specification: 1st Stg OLS IV Reduced Form

Panel A: Low Level in Hierarchy
MOR in other period 0.349*** -0.124***
(0.037) (0.043)
MOR in current period -0.164*** -0.354***
(0.043) (0.127)
Mean dep. var. 1.532 1.532 1.532
F-stat on excl instrument 86.97
Panel B: Medium Level in Hierarchy
MOR in other period 0.274*** -0.190***
(0.037) (0.045)
MOR in current period -0.145*** -0.691***
(0.045) (0.180)
Mean dep. var. 1.103 1.103 1.103
F-stat on excl instrument 54.12
Panel C: High Level in Hierarchy
MOR in other period 0.198*** -0.248**
(0.075) (0.111)
MOR in current period -0.136 -1.253*
(0.092) (0.751)
Mean dep. var. 1.425 1.425 1.425
F-stat on excl instrument 7.039

Notes: Each panel is similar to Panel A of Table 3. The difference is that we examine heterogeneity by
heterogeneity in the firm hierarchy. As discussed in Section A.8, we divide individuals at the firm into three
levels of hierarchy according to their salary grade, following how the firm often segments employees in its
internal reporting. * significant at 10%; ** significant at 5%; *** significant at 1%

31
 Table C19: The Standard Deviation of Manager Value-added

Method: One Sample Split Sample Split Sample

(split (split
randomly) by period)
(1) (2) (3)
Panel A: Attrition (x100)
SD of mgr effects 1.20 0.67 0.67
Panel B: Quits (x100)
SD of mgr effects 0.82 0.38 0.48
Panel C: Fires (x100)
SD of mgr effects 0.55 0.33 0.19
Panel D: Regretted Quits (x100)
SD of mgr effects 0.71 0.29 0.40
Panel E: Non-regretted Quits (x100)
SD of mgr effects 0.35 0.13 0.08
Notes: This table presents estimates of the standard deviation of manager value-added for five attrition
variables (attrition, quits, fires, regretted quits, and non-regretted quits). In all columns, we estimate a
version of equation (7) from the main text while using the baseline controls from Table 3 (excluding MOR).
We use the same data as from Table 3 where an observation is an employee-month. The standard deviations
shown are weighted by the number of observations. In (1), we estimate one set of manager fixed effects using
the full sample. In (2), we randomly split the data in two, randomly assigning each employee-month to one
of two samples. The standard deviations shown are calculated using the covariance of the fixed effects
estimated using the two samples. (The standard deviations shown are the square root of the estimated
covariances.) In (3), we split the data into the first and second periods, and estimate manager fixed effects
separately by period. The standard deviations in (2) and (3) are smaller than in (1) as they are adjusted for
sampling error.

32
 Table C20: MOR and Non-Attrition Outcomes: Exploiting New Joiners

Dep. Var.: Subjective Log Salary Promotion

performance Growth
(x100)
(1) (2) (3)
Panel A: OLS
MOR (normalized) 0.032 -0.003 -0.023
(0.026) (0.319) (0.036)
Panel B: IV
MOR (normalized) -0.060 -0.729 -0.047
(0.084) (0.904) (0.152)
Panel C: Reduced Form
MOR (normalized) -0.018 -0.200 -0.014
(0.024) (0.246) (0.045)
33

Employee FE No No No
Notes: Standard errors clustered by manager in parentheses. The specifications are similar to the odd columns in Table 8, but this table restricts to
new employees joining the firm after the administration of the second survey (as in Table 4). * significant at 10%; ** significant at 5%; ***significant
at 1%
 Table C21: MOR and Non-Attrition Outcomes: Exploiting New Joiners and People Switching Managers

Dep. Var.: Subjective Log Salary Promotion

performance Growth
(x100)
(1) (2) (3)
Panel A: OLS
MOR (normalized) 0.027** 0.210 0.191***
(0.011) (0.171) (0.053)
Panel B: IV
MOR (normalized) -0.014 -0.472 0.217
(0.048) (0.554) (0.272)
Panel C: Reduced Form
MOR (normalized) -0.004 -0.149 0.061
(0.013) (0.176 (0.076)
34

Employee FE No No No
Notes: Standard errors clustered by manager in parentheses. The specifications are similar to the odd columns in Table 8, but restricts to employees
who experience their first change (during our data period) in manager (as in Table 5). * significant at 10%; ** significant at 5%; ***significant at 1%
 Table C22: MOR and Employee Engagement

(1) (2)
Panel A: OLS
MOR (normalized) 0.048*** 0.027
(0.011) (0.017)
Panel B: IV
MOR (normalized) 0.069** 0.018
(0.034) (0.026)
Panel C: Red. Form
MOR (normalized) 0.023** -0.009
(0.011) (0.017)
Employee FE No Yes
Notes: Standard errors clustered by manager in parentheses. The specifications are similar in controls and
format to those in Table 8. The difference is that the dependent variable here is normalized employee
engagement. * significant at 10%; ** significant at 5%; ***significant at 1%

35
 Table C23: Robustness for Significant Results in Table 9: Gradually Adding Controls

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

Panel A: Subjective Perf
MOR in current period 0.419*** 0.447*** 0.470*** 0.510*** 0.506***
(0.0870) (0.0904) (0.0933) (0.102) (0.101)
Panel B: Promotions
MOR in current period 0.673** 0.726** 0.765** 0.739** 0.717**
(0.311) (0.325) (0.341) (0.365) (0.360)
Panel C: Log Salary Growth (x100)
MOR in current period 1.405** 1.633** 1.935*** 2.283*** 2.298***
(0.627) (0.653) (0.733) (0.882) (0.880)
Base Controls Y Y Y Y Y
Business Unit X Job Function Dummies N Y Y Y Y
Business Unit X Salary Grade Dummies N N Y Y Y
36

Job Function X Salary Grade Dummies N N N Y Y

Current Month Dummies N N N N Y
Notes: Standard errors clustered by manager in parentheses. Column 1 here is the same as column 1 (Panels A and B) of Table 9. Columns 2-5
subsequently add additional controls, similar to Tables C12-C14. * significant at 10%; ** significant at 5%; *** significant at 1%
 Table C24: What are Managers Rewarded For? Employees Survey Scores (MOR), Period 1 Only

Dep var: Subjective Promoted Log Log Log Log Change in Key
performance (x100) salary salary stock change in span of individual
(normalized) (x100) growth grant stock control (x100)
(x100) holdings grants
(x100) (x100)
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A: OLS
MOR in current period 0.102*** 0.00172 -0.867 0.248 0.0852 -2.592 0.111 -2.289
(0.0378) (0.221) (0.640) (0.218) (2.299) (5.378) (0.135) (1.723)
Panel B: IV
MOR in current period 0.195** 0.936 -0.223 0.518 -3.626 22.66 0.196 2.879
(0.0926) (0.573) (1.644) (0.582) (6.289) (14.36) (0.357) (4.292)
Panel C: Red. Form
37

MOR in other period 0.0725** 0.348* -0.0832 0.195 -1.311 7.890 0.0732 1.072
(0.0340) (0.207) (0.621) (0.222) (2.267) (4.885) (0.134) (1.589)

Notes: Standard errors clustered by manager in parentheses. This table is similar to Table 9 in the main text except that we restrict attention to
manager-months using the first period of the data (i.e., the data on or before September Y1 ). * significant at 10%; ** significant at 5%; *** significant
at 1%
 Table C25: What are Managers Rewarded For? Employees Survey Scores vs. VA

Dep var: Subjective Promoted Log Log Log Log Change in Key
performance (x100) salary salary stock change in span of individual
(normalized) (x100) growth grant stock control (x100)
(x100) holdings grants
(x100) (x100)
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A: OLS
MOR in current period 0.0611*** 0.0950 -0.471 0.120 -2.657* -0.169 0.0672 -1.104
(0.0226) (0.0903) (0.410) (0.167) (1.558) (3.229) (0.0941) (0.864)
Manager FE in retention 0.0379 0.0774 -0.0753 0.152 -0.328 4.620 -0.0417 2.764***
(0.0232) (0.0796) (0.536) (0.209) (2.003) (3.275) (0.108) (0.961)
Panel B: IV
38

MOR in current period 0.410*** 0.636 -2.438 1.995* -2.631 0.306 0.339 0.953
(0.128) (0.418) (2.818) (1.021) (7.459) (14.49) (0.411) (4.091)
Manager FE in retention 0.0385 0.153 0.156 -1.898 7.662 19.75 -0.365 4.541
(0.291) (0.885) (5.714) (2.077) (16.48) (27.26) (1.132) (9.759)
Panel C: Red. Form
MOR in other period 0.134*** 0.214** -0.719 0.474** -0.286 1.571 0.0861 0.620
(0.0223) (0.0947) (0.450) (0.192) (1.702) (3.187) (0.0983) (0.882)
Manager FE in retention 0.0434 0.0801 -0.212 -0.0980 0.903 3.308 -0.0184 0.701
(0.0292) (0.0929) (0.577) (0.196) (2.033) (3.321) (0.108) (1.034)

Notes: Standard errors clustered by manager in parentheses. This table is similar to Table 9. The difference is that we also include the manager’s
fixed effect in retention (i.e., the manager’s value added or VA) as a regressor. In calculating the manager’s fixed effect, we randomly split the sample
of all worker-months in two. Using this, we estimate two manager fixed effects for each manager, hereafter M0 and M1 . In the OLS, we use M0 as the
regressor. In the IV, we use M0 as the regressor and use M1 as an instrument. In the reduced form, we use M1 as the regressor. For a very small
number of observations (8 person-months), M1 is missing; for these observation, we mean-impute the missing values (i.e., we assign the fixed effect to
0). * significant at 10%; ** significant at 5%; *** significant at 1%
 Table C26: What are Managers Rewarded For? Employees Survey Scores vs. Subjective Performance Score vs. VA

Dep var: Subjective Promoted Log Log Log Log Change in Key
performance (x100) salary salary stock change in span of individual
(normalized) (x100) growth grant stock control (x100)
(x100) holdings grants
(x100) (x100)
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A: OLS
MOR in current period 0.0611*** -0.00709 -0.540 0.0261 -3.292** -1.373 0.0764 -1.555*
(0.0226) (0.0909) (0.412) (0.167) (1.520) (3.201) (0.0967) (0.843)
Subj performance 1.701*** 0.853*** 0.904*** 11.68*** 11.39*** -0.0995 6.229***
(0.111) (0.327) (0.120) (1.246) (2.308) (0.0958) (0.739)
Manager FE in retention 0.0379 0.00595 -0.124 0.131 -0.702 4.801 -0.0662 2.636***
(0.0232) (0.0798) (0.528) (0.211) (1.973) (3.323) (0.109) (0.974)
Panel B: IV
39

MOR in current period 0.410*** -0.0370 -3.154 1.638* -6.360 -4.123 0.219 -1.496
(0.128) (0.367) (2.589) (0.903) (6.972) (13.78) (0.391) (3.812)
Subj performance 1.700*** 0.968*** 0.831*** 11.70*** 11.18*** -0.107 6.154***
(0.112) (0.356) (0.145) (1.295) (2.380) (0.0970) (0.767)
Manager FE in retention 0.0385 0.0667 0.631 -1.830 5.829 20.31 -0.148 4.553
(0.291) (0.752) (5.070) (1.718) (15.04) (25.02) (1.009) (9.116)
Panel C: Red. Form
MOR in other period 0.134*** -0.00743 -0.891** 0.370* -1.601 0.132 0.0617 -0.178
(0.0223) (0.0921) (0.449) (0.192) (1.671) (3.169) (0.100) (0.871)
Subj performance 1.701*** 0.927*** 0.886*** 11.64*** 11.29*** -0.102 6.235***
(0.111) (0.329) (0.120) (1.273) (2.332) (0.0960) (0.754)
Manager FE in retention 0.0434 0.00596 -0.201 -0.150 0.310 3.011 -0.00320 0.501
(0.0292) (0.0868) (0.566) (0.190) (2.034) (3.224) (0.111) (1.030)

Notes: Standard errors clustered by manager in parentheses. This table is similar to Table C25. The difference is that we further add a manager’s
own subjective performance score (received from his/her superiors) as a regressor. * significant at 10%; ** significant at 5%; *** significant at 1%
Appendix References
Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber, “Selection on Observed
and Unobserved Variables: Assessing the Effectiveness of Catholic Schools,” Journal of
Political Economy, 2005, 113 (1), 151–184.

Angrist, Joshua D and Jörn-Steffen Pischke, Mostly Harmless Econometrics: An Em-

piricist’s Companion, Princeton university press, 2008.

Bloom, Nicholas, Renata Lemos, Raffaella Sadun, Daniela Scur, and John Van
Reenen, “JEEA-FBBVA Lecture 2013: The New Empirical Economics of Management,”
Journal of the European Economic Association, 2014, 12 (4), 835–876.

Chetty, Raj, John N. Friedman, and Jonah E. Rockoff, “Measuring the Impacts
of Teachers I: Evaluating Bias in Teacher Value-Added Estimates,” American Economic
Review, September 2014, 104 (9), 2593–2632.

Enikolopov, Ruben, Alexey Makarin, and Maria Petrova, “Social Media and Protest
Participation: Evidence from Russia,” 2017. Working paper, Universitat Pompeu Fabra.

Garicano, Luis, “Hierarchies and the Organization of Knowledge in Production,” Journal

of Political Economy, 2000, 108 (5), 874–904.

Hoffman, Mitchell, Matthew Bidwell, John McCarthy, and Michael Housman,

“The Determinants of Managerial Productivity around the World,” 2018. Mimeo, University
of Toronto.

Oster, Emily, “Unobservable Selection and Coefficient Stability: Theory and Evidence,”
Journal of Business & Economic Statistics, 2017, 0 (0), 1–18.

Satyanath, Shanker, Nico Voigtländer, and Hans-Joachim Voth, “Bowling for fas-
cism: social capital and the rise of the Nazi Party,” Journal of Political Economy, 2017,
125 (2), 478–526.

40