DOES MANAGEMENT MATTER?

EVIDENCE FROM INDIA

Nicholas Blooma , Benn Eifertb, Aprajit Mahajanc,

David McKenzied and John Robertse

Preliminary draft: October 18, 2010

Abstract:
A long-standing question in social science is the extent to which differences in management cause
differences in firm performance. To investigate this we ran a management field experiment on large
Indian textile firms. We provided free consulting on modern management practices to a randomly
chosen set of treatment plants and compared their performance to the control plants. We find that
adopting modern management practices had three main effects. First, it raised average productivity
by 11%, through improved quality and efficiency and reduced inventory. Second, it increased the
decentralization of decision making, as the better flow of information enabled owners to delegate
more decisions to plant managers. Third, it increased the use of computers, necessitated by the data
collection and analysis involved in modern management. Since these practices were profitable this
raises the question of why firms had not adopted these before. Our results suggest that informational
barriers were a primary factor in explaining this lack of adoption. Modern management is a type of
technology that diffuses slowly between firms, with many Indian firms unaware of its existence or
impact. And since competition was limited by constraints on firm entry and growth, badly managed
firms were not rapidly driven from the market.

JEL No. L2, M2, O14, O32, O33.

Keywords: management, organization, IT, productivity and India.
Acknowledgements: We are deeply thankful to Benn Eifert, who was instrumental in getting this
project underway. Financial support was kindly provided by the Alfred Sloan Foundation, the
Freeman Spogli Institute, the International Initiative at Stanford, the Graduate School of Business at
Stanford, the International Growth Centre, IRISS, the Kauffman Foundation, the Murthy Family,
the Knowledge for Change Trust Fund, the National Science Foundation, the Toulouse Network for
Information Technology, and the World Bank. This research would not have been possible without
our partnership with Kay Adams, James Benton and Breck Marshall, the dedicated work of the
consulting team of Asif Abbas, Saurabh Bhatnagar, Shaleen Chavda, Karl Gheewalla, Shruti
Rangarajan, Jitendra Satpute, Shreyan Sarkar, and Ashutosh Tyagi, and the research support of Troy
Smith. We thank our formal discussants Susantu Basu, Naushad Forbes, Ramada Nada and Paul
Romer, as well as seminar audiences at the AEA, Barcelona GSE, Boston University, Chicago,
Columbia, the EBRD, Harvard Business School, IESE, Katholieke Universiteit Leuven, Kellogg,
the LSE, the NBER, PACDEV, Stanford, TNIT, UCL, UCLA, Wharton and the World Bank for
comments.
a
Stanford Economics, SCID, CEP and NBER; b Berkeley Economics; c
Stanford Economics and
SCID; d The World Bank, IZA and BREAD; e Stanford GSB

1
 I. INTRODUCTION

Economists have long puzzled over why there are such astounding differences in productivity
across both firms and countries. For example, US plants in homogeneous industries like cement,
block-ice, white pan bread and oak flooring display 100% productivity spreads between the 10th and
90th percentile (Foster, Haltiwanger and Syverson, 2008).

A natural explanation for these productivity differences lies in variations in management practices.
Indeed, the idea that “managerial technology” determines the productivity of inputs goes back at
least to Walker (1887) and is central to the Lucas (1978) model of firm size. Yet while management
has long been emphasized by the media, business schools and policymakers, economists have
typically been skeptical about its importance.

One reason for skepticism is the belief that competition will drive badly managed firms out of the
market. As a result any residual variations in management practices will reflect firms’ optimal
responses to differing market conditions. For example, firms in developing countries may not be
adopting quality control systems because wages are so low that repairing defects is cheap. Hence,
their management practices are not “bad”, but the optimal response to low wages.

A second reason for this skepticism is the complexity of management, making it hard to measure
and quantify.1 However, recent work has down-played the “soft skill” attributes of good managers –
which can be difficult to measure, let alone change – in order to focus on specific management
practices which can be measured, taught in business schools and recommended by consultants.
Examples of such practices include key principles of Toyota’s “lean manufacturing,” the continual
analysis and refinement of quality control procedures, inventory management, and advanced human
resource practices. A growing literature measures many such practices and finds large variations
across establishments and a strong association between these management practices and higher
productivity and profitability.2

This paper seeks to provide the first experimental evidence of the importance of management
practices in large firms. The experiment takes large multi-plant Indian textile firms and randomly
allocates their plants to management treatment and control groups. Treatment plants received five
months of extensive management consulting from a large international consulting firm, which
diagnosed areas for improvement in a set of management practices in the first month, followed by
four months of intensive support in implementation of these recommendations. The control plants
received only the one month of diagnostic consulting.

The treatment intervention led to significant improvements in quality, inventory and production
efficiency (defined as the fraction of time the weaving machines were running) The result was an
increase in productivity of 11% and a substantial increase in annual profitability of about $230,000.

1
Lucas (1978, p. 511) notes that his model “does not say anything about the tasks performed by managers, other than
whatever managers do, some do it better than others”.
2
See for example, Osterman (1994), Huselid (1995), MacDuffie (1995), Ichniowski, Prennushi and Shaw (1998),
Cappelli and Neumark (2001) and Bloom and Van Reenen (2007). A prominent early example is Pack (1987), which,
like the present study, deals with textile firms in developing countries. In related work, Bertrand and Schoar (2003) use
a manager-firm matched panel and find that manager fixed effects matter for a range of corporate decisions. Lazear and
Oyer (2009) and Bloom and Van Reenen (2010) provide extensive surveys.

2
Firms also spread these management improvements from their treatment plants to other plants they
owned, providing revealed preference evidence on their beneficial impact.

The improvements were substantial because our sample of plants had very poor management
practices prior to the intervention. Most of them had not adopted basic procedures for efficiency,
inventory or quality control that have been commonly used for several decades in similarly sized
European, US and Japanese firms. Since this consulting would have cost the firms about $250,000
if they had paid for this – implying about an 90% rate of return given the $230,000 increase in
profits - this raises the question of why these practices had not been adopted before?

Our evidence suggests that one important factor was informational constraints – the Indian firms
were not aware of the existence or impact of common modern management practices. This is
perhaps not entirely surprising. Management practices evolve gradually over time, with innovations
like the Taylor's Scientific Management, Sloan's M-form corporation and Toyota's lean production
spreading slowly across firms and countries. For example, the US automotive industry took at least
two decades to understand and adopt Japanese lean manufacturing.

A related question is why product market competition does not drive these badly managed firms out
of business. One reason is the reallocation of market share to well managed firms is restricted by
span of control constraints on firm growth. In every firm in our sample all senior managerial
positions are held by members of the owning family. The number of adult males available to fill
senior positions thus becomes a binding constraint on growth.3 Hence, well managed firms do not
always grow large and drive unproductive firms out of the market if they lack male family
members.4 Meanwhile, entry is limited by financing costs (our textile firms have assets of $13m on
average), while imports are restricted by heavy tariffs.

We also find two other major results for the impact of better management practices. First, owners
decentralized greater decision making power over hiring, investment and pay to their plant
managers. This happened in part because the improved collection and dissemination of information
enabled owners to monitor their plant managers better, reducing the risk of managerial theft. As a
result owners felt more comfortable in delegating decision making power to middle managers.

Second, the extensive data collection and processing requirements of modern management led to a
rapid increase in computer use. For example, installing production quality control systems requires
firms to record individual quality defects and then to analyze these by shift, loom, weaver and
design. So modern management practices appear to be a skill-biased technical change (SBTC) as
increased computerization raises the demand for educated employees. A large literature has
highlighted SBTC as a key factor increasing income inequality since the 1970s, with this
experiment providing some evidence on the role of management as one mechanism.5

The major drawback of our experiment is the small cross-sectional sample size. We have data on
only 28 plants across 17 firms. To address concerns over statistical inference in small samples we
implement permutations tests that have exact finite sample size. We also exploit our large time

3
The best predictor of firm size was the number of male family members. All the biggest firms had multiple plants run
by multiple brothers, while the best managed firm had only one plant and the founder had no brothers or sons.
4
This may also help to explain the lack of reallocation across firms in China and India (Hsieh and Klenow, 2009).
5
See, for example, the surveys in Acemoglu 2002 and Autor, Katz and Kearney 2008.

3
series of around 100 weeks of data per plant by using robust estimators that rely on large T (rather
than large N) asymptotics. We believe these approaches are useful both for addressing sample
concerns in our paper and also potentially for other field experiments on large organizations,
regions or villages where the data has a small cross-section but large time series.

This paper relates to several strands of literature. First, there is the extensive productivity literature
which reports large spreads in total-factor productivity (TFP) across plants and firms in dozens of
developed countries. From the outset this literature has attributed much of this spread to differences
in management practices (Mundlak, 1961), but problems in measurement and identification have
made this hard to confirm (Syversson, 2010). This dispersion in productivity appears even larger in
developing countries (Banerjee and Duflo, 2005, Hsieh and Klenow, 2009). Despite this, there are
still very few experiments on productivity in firms (McKenzie, 2009) and none involving large
multi-plant firms.

Second, our paper builds on the literature on the management practices of firms. This has a long
debate between the “best-practice” view that some management practices are universally good and
all firms would benefit from adopting these (Taylor, 1911) and the “contingency view” that every
firm is already adopting optimal practices but these are different from firm to firm (e.g. Woodward,
1958). Much of the empirical literature trying to distinguish between these views has traditionally
been case-study or survey based, making it hard to distinguish between the different explanations
and resulting in little consensus in the management literature.6 This paper provides experimental
evidence that a core set of management best practices do exist.

Third, the paper links to the large theoretical literature on the organization of firms. Papers
generally emphasize optimal decentralization either as a way to minimize information processing
costs or as a way to trade off incentives and information within a principal-agent model.7 But the
empirical evidence on delegation is limited, focusing on natural experiments like the adoption of
on-board computers in trucking (e.g. Baker and Hubbard, 2004) or de-layering in large publicly
traded US firms (Rajan and Wulf, 2006, Guadalupe and Wulf, 2010). In this paper we present the
first experimental evidence on decentralization.

Fourth, it links with the rapidly growing literature on Information Technology (IT) and productivity.
A growing body of work has emphasized the relationship between technology and productivity,
emphasizing both the direct productivity impact of IT and also its complementarity with modern
management and organizational practices (i.e. Bresnahan et al. 2002 and Bartel, Ichniowski and
Shaw, 2007). But again the evidence on has focused on observational IT and organizational survey
data, with no prior experimental data. Our experimental evidence suggests one route for the impact
of computers on productivity is via facilitating better management practices, and this occurs
simultaneously with the decentralization of production decisions.

Finally, recently a number of other field experiments in developing countries (for example Karlan
and Valdivia, 2010, Bruhn et al. 2010 and Drexler et al. 2010) have begun to estimate the impact of

6
See, for example, the surveys in Delery and Doty (1996) and Bloom and Van Reenen (2010).
7
See, for example, Bolton and Dewatripont (1994) and Garicano (2000) for examples of information processing models
and Aghion and Tirole (1997), Baker, Gibbons and Murphy (1999), Rajan and Zingales (2001), Hart and Moore (2005),
Acemoglu et al. (2007) and Alonso et al. (2008) for examples of principal-agent models. Recent reviews of this
literature are contained in Garicano and Van Zandt (2010), Mookherjee (2010) and Gibbons and Roberts (2010).

4
business training in micro and small enterprises. This work focuses on training the owners in tasks
such as separating business and personal finances, basic accounting, marketing and pricing. This
research generally finds significant effects of these business skills on performance in small firms,
supporting our results on management practices in larger firms.

II. MANAGEMENT IN THE INDIAN TEXTILE INDUSTRY

II.A. Why work with firms in the Indian textile industry?

Despite rapid growth over the past decade, India’s one billion people still have labor productivity of
only 15 percent of that in the U.S. (McKinsey Global Institute, 2001). While average levels of
productivity are low, most notable is the large variation in productivity, with a few highly
productive firms and a lot of low-productivity firms (Hsieh and Klenow, 2009).

In common with other developing countries for which data is available, Indian firms are also
typically poorly managed. Evidence from this is seen in Figure 1, which plots results from the
Bloom and Van Reenen (2010) surveys of manufacturing firms in the US and India. The Bloom and
Van Reenen (BVR) methodology scores establishments from 1 (worst practices) to 5 (best
practices) on specific management practices related to monitoring, targets and incentives. This
yields a basic measure of the use of modern management practices and is strongly correlated with a
wide range of firm performance measures like productivity, profitability and growth. The top panel
of Figure 1 plots the these management practice scores for a sample of 751 randomly chosen US
manufacturing firms with 100 to 5000 employees and the second panel for similarly sized Indian
ones. The results reveal a thick tail of badly run Indian firms, leading to a much lower average
management score (2.69 for India versus 3.33 for US firms). Indian firms tend not to collect and
analyze data systematically in their factories, they tend to use less effective target-setting and
monitoring, and they typically employ ineffective promotion and reward systems. The scores for
Brazil and China in the third panel, with an average of 2.67, are very similar, suggesting that Indian
firms are representative of large manufacturing firms in the emerging economies.

In order to implement a common set of management practices across firms and measure a common
set of outcomes, we focus on one industry. We chose textile production since it is the largest
manufacturing industry in India, accounting for 22% of manufacturing employment. The fourth
panel shows the management scores for the 232 textile firms in the BVR Indian sample, which look
very similar to Indian manufacturing in general.

Within textiles, our experiment was carried out on 28 plants operated by 17 firms in the woven
cotton fabric industry. These plants weave cotton yarn into cotton fabric for suits, shirting and home
furnishing. They purchase yarn from upstream spinning firms and send their fabric to downstream
dyeing and processing firms. As shown in the bottom panel of Figure 1, the 17 firms involved had
an average BVR management score of 2.60, very similar to the rest of Indian manufacturing. Hence,
our particular sample of 17 Indian firms also appears broadly similar in terms of management
practices to manufacturing firms in developing countries.8

8
Interestingly, prior work on the Indian textile industry suggested its management practices were also inferior to those
in Europe in the early 1900s (Clark, 1987).

5
II.B. The selection of firms for the field experiment

The sample firms were randomly chosen from the population of all publicly and privately owned
textile firms in Maharashtra, based on lists provided to us by the Ministry of Corporate Affairs
(MCA). We restricted attention to firms with between 100 to 1000 employees.9 We chose 100
employees as the lower threshold because by this size firms require systematic management
practices to operate efficiently. We chose 1000 employees as the upper bound to avoid working
with conglomerates and multinationals. Geographically we focused on firms in the towns of
Tarapur and Umbergaon because these provide the largest concentrations of textile firms in the area,
and concentrating on two nearby towns reduced travel time for the consultants. This yielded a
sample of 66 potential subject firms.

All of these 66 firms were then contacted by telephone by our partnering international consulting
firm. They offered free consulting, funded by Stanford University and the World Bank, as part of a
management research project. We paid for the consulting to ensure we controlled the intervention
and could provide a homogeneous management treatment to all firms. We were concerned that if
the firms made any co-payments they might have tried to direct the consulting, for example asking
for help on marketing or finance.

Of this group of firms, 34 expressed an interest in the project and were given a follow-up visit and
sent a personally signed letter from Stanford. Of the 34 firms, 17 agreed to commit senior
management time to the consulting program.10 We compared these program firms with the 49 non-
program firms and found no significant differences in observables.11

The experimental firms have typically been in operation for 20 years and all are family-owned.
They produce fabric for the domestic market, with many firms also exporting. Table 1 reports some
summary statistics for the textile manufacturing parts of these firms (many of the firms have other
businesses in textile processing, retail and real estate). On average these firms had about 270
employees, current assets of $13 million and sales of $7.5m a year. Compared to US manufacturing
firms these firms would be in the top 2% by employment and the top 5% by sales,12 and compared
to India manufacturing in the top 1% by both employment and sales (Hsieh and Klenow, 2010).
Hence, by this criterion, as well as by most formal definitions, these are large manufacturing
firms.13

9
The MCA list comes from the Registrar of Business, with whom all public and private firms are legally required to
register annually. Of course many firms do not register in India, but this is generally a problem with smaller firms, not
with 100+ employee manufacturing firms which are too large and permanent to avoid Government detection.
10
The two main reasons for refusing free consulting given on the telephone and during the visits was that the firms did
not believe they needed management assistance or that it required too much time from their senior management (1 day a
week). But it is also possible the real reason is these firms were suspicious of this offer, given many firms in India have
tax and regulatory irregularities.
11
For example, the program firms had slightly less assets ($12.8m) compared to the non-program firms ($13.9m), but
this difference was not statistically significant (p-value 0.841). We also compared the groups on management practices
using the BVR scores, and found they were almost identical (difference of 0.031, p-value 0.859).
12
Dunn & Bradstreet (August 2009) lists 778,000 manufacturing firms in the US with only 17,300 of these (2.2%) with
270 or more employees and only 28,900 (3.7%) with $7.5m or more sales.
13
Most international agencies define large firms as those with more than 250+ employees.

6
These firms are complex organizations, with a median of 2 plants per firm and 4 reporting levels
from the shop-floor to the managing director. In all the firms, the managing director is the single-
largest shareholder, reflecting the lack of separation of ownership and management in many Indian
firms. All directors are family members. One firm is publicly quoted on the Mumbai Stock
Exchange, although more than 50% of the equity is held by the managing director and his father.

In exhibits (1) to (7) we include a set of photographs of the plants. These are included to provide
some background information to readers on their size, production process and initial state of
management. As is clear these are large establishments (Exhibit 1), with several multi-story
buildings per site, and typically two production sites per firm, plus a head office in Mumbai. They
operate a continuous production process that runs constantly (Exhibit 2). Their factories’ floors
were (initially) often rather disorganized (Exhibits 3 and 4), and their yarn and spare-parts inventory
stores lacking any formalized storage systems (Exhibits 5 and 6).

III. THE MANAGEMENT INTERVENTION

III.A. Why use management consulting as an intervention

The field experiment aimed to improve management practices in the treatment plants. To achieve
this we hired a management consultancy firm to work with our treatment plants as the operationally
easiest way to rapidly change plant level management. We selected the consulting firm using an
open tender. The winner was a large international management, IT and outsourcing consultancy
which is headquartered in the U.S., but has about 40,000 employees in India. The full-time team of
(up to) 6 consultants working on the project at any time all came from their Mumbai office. These
consultants were educated at leading Indian business and engineering schools, and most of them had
prior experience working with US and European multinationals.

Selecting a high profile international consulting firm substantially increased the cost of the project.
But it meant that our experimental firms were more prepared to trust the consultants, which was
important for getting a representative sample group. It also offered the largest potential to improve
the management practices of the firms in our study.

The project ran from August 2008 until August 2010, and the total cost of this was US$1.3 million,
or approximately $75,000 per treatment plant and $20,000 per control plant.14 Note this is very
different from what the firms themselves would pay for comparable consulting, which would be
probably about $250,000. The reason for our much cheaper costs per plant is that, because it was a
research project, Accenture charged us pro-bono rates (50% of commercial rates) and provided
partners’ time at no cost. In addition, it realized substantial economies of scale by working across
multiple plants.

While the intervention offered high-quality management consulting services, the purpose of our
study was to use the improvements in management generated by this intervention to understand if

14
These costs may seem high for India. Certainly, at the bottom of the consulting quality distribution in India
consultants are cheap, but their quality is poor. At the top end, rates are more comparable to those in the US. This is
because international consulting companies target multinationals and employ consultants that are often US or European
educated and have access to international labor markets.

7
(and how) modern management practices affect firm performance. Like many recent development
projects, this intervention was provided as a mechanism of convenience – to change management
practices – and not to evaluate the management consultants themselves.

III.B. The management consulting intervention

The intervention aimed to introduce a set of standard management practices. Based on their prior
textile experience, the consultants identified 38 key practices on which to focus. These 38 practices
encompass a range of basic manufacturing principles that are standard in almost all US, European
and Japanese firms, and can be grouped into five areas:

• Factory Operations and Planning: Plants were encouraged to undertake regular maintenance of
machines (rather than repairing machines only when they broke down) and record the reasons
for machine downtime so they could learn from past failures. They were encouraged to keep
the factory floor tidy to reduce accidents and ease the movement of materials.

• Quality control: Plants were encouraged to record quality defects by type at every stage of the
production process, analyze these records daily, and formalize procedures to address defects to
prevent them occurring repeatedly.

• Inventory: Plants were encouraged to record yarn stocks, on a daily basis, with optimal
inventory levels being defined and stock monitored against these. Yarn was to be sorted,
labeled and stored in the warehouse by type and color, and this information logged onto a
computer so yarn could be located when required for production.

• Human-resource management: Plants were encouraged to introduce a performance-based

incentive system for workers and managers. Worker incentives were to be linked to output,
quality and attendance. Job descriptions were to be defined for all workers and managers.

• Sales and order management: Plants were encouraged to track production on an order-wise
basis to prioritize customer orders by delivery deadline. Design-wise efficiency analysis was
suggested so that pricing could be based on design (rather than average) production costs.

These 38 management practices (listed in Appendix Table A1) form a set of precisely defined
binary indicators that we can use to measure improvements in management practices as a result of
the consulting intervention.15 We recorded these indicators on an on-going basis throughout the
study. A general pattern at baseline was that plants recorded a variety of information (often in paper
sheets), but had no systems in place to monitor these records or use them in decisions. Thus, while
93 percent of the treatment plants recorded quality defects before the intervention, only 29 percent
monitored them on a daily basis or by the particular sort of defect, and none of them had any
standardized analysis and action plan based on this defect data.

15
We prefer these indicators to the BVR management score for our work here, since they are all binary indicators of
specific practices, which are directly linked to the intervention. In contrast, the BVR indicator measures practices at a
more general level on a 5-point ordinal scale. Nonetheless, the sum of our 38 pre-intervention management practice
scores is correlated with the BVR score at 0.404 (p-value of 0.077) across the 17 firms.

8
The consulting treatment had three stages. The first stage, called the diagnostic phase, took one
month and was given to all treatment and control plants. It involved evaluating the current
management practices of each plant and constructing a performance database. Construction of this
database involved setting up processes for measuring a range of plant-level metrics – such as
output, efficiency, quality, inventory and energy use – on an ongoing basis, plus developing and
recording historical data from existing plant records. For example, to facilitate quality monitoring
on a daily basis, a single metric constructed as a severity-weighted average of the major types of
defects, termed the Quality Defects Index (QDI) was defined.

At the end of the diagnostic phase the consulting firm provided each treatment and control plant
with a detailed analysis of its current management practices and performance. The treatment plants
were given this diagnostic phase as the first step in improving their management practices. The
control plants were given it only because we needed to construct historical performance data for
them and help set up systems to generate ongoing data.

The second phase was a four month implementation phase which was given only to the treatment
plants. In this the consulting firm followed up on the diagnostic report to help introduce as many of
the 38 key management practices as the firms could be persuaded to adopt. The consultant assigned
to each plant worked with the plant management to put the procedures into place, fine-tune them,
and stabilize them so that they could be readily carried out by employees. For example, one of the
practices implemented was daily meetings for management to review production and quality data.
The consultant attended these meetings for the first few weeks of the implementation phase to help
the managers run them, provided feedback on how to run future meetings, and fine-tuned their
design to the specific plant’s needs.

The third phase was a measurement phase which lasted until August 2010. This phase involved only
three consultants and a part-time manager, who collected performance and management data from
the plants. In return for continuing to provide this data the consultants continued to provide some
light consulting advice to both the treatment and control plants.

So, in summary, the control plants were provided with just the diagnostic phase and then the
measurement phase (totaling 225 consultant hours on average), while the treatment plants were
provided with the diagnostic, then implementation and then measurement phases (totaling 733
consultant hours on average).

III.C. The experimental design

We wanted to work with large firms because their complexity means management practices are
likely to be important. However, providing consulting to large firms is expensive, which
necessitated a number of trade-offs which are detailed below.

Sample size:
We worked with the 28 plants within our 17 experimental firms. We considered hiring cheaper local
consultants and providing more limited consulting to a sample of several hundred plants. But two
factors pushed against this. First, many large firms in India are reluctant to let outsiders into their
plants because of their lack of compliance with tax, labor and safety regulations. To minimize
selection bias we offered a high quality intensive consulting intervention that firms would value
enough to take the risk of allowing outsiders into their plants. This helped maximize initial take-up

9
(26% as noted in section II.B) and retention (100%, as no firms dropped out). Second, the
consensus from discussions with Indian business people was that achieving a measurable impact in
large firms would require an extended engagement with high-quality consultants. Obviously the
trade-off was this led to a small sample size,. We discuss the estimation issues this generates in
section III.D below.

On-site and off-site plants: Due to manpower constraints we could collect detailed performance data
from only 20 plants. Building data collection systems and compiling historic databases required the
consultants spending several hours each week on-site at each plant. As a result the performance
regressions are run only on these 20 “experimental” plants. However, data on management,
organizational and IT outcomes were gathered for all 28 plants, as this required only bi-monthly
visits to the other eight. These eight plants are referred to as “non-experimental”.

Treatment and control plants: Within the group of 20 experimental plants we randomly picked 6
control plants and 14 treatment plants. As Table 1 shows, the treatment and control firms were not
statistically different across any of the characteristics we could observe.16

Timing: The consulting intervention was phased in three waves because of the capacity constraint
of the six person consulting team. So the first wave started in September 2008 with 4 treatment
plants. In April 2009 a second wave of 10 treatment plants was initiated, and in July 2009 the third
wave of 6 control plants was initiated. Firm records allow us to collect data going back to a
common starting point of April 2009.

We started with a small first wave because we expected the intervention process to get easier over
time due to accumulated experience. The second wave included all the remaining treatment firms
because: (i) the consulting interventions take time to affect performance and we wanted the longest
time-window to observe the treatment firms; and (ii) we could not mix the treatment and control
firms across waves.17 The third wave contained the control firms. We picked more treatment than
control plants because the staggered initiation of the interventions meant the different treatment
groups provided some cross identification for each other, and because we believed the treatment
plants would be more useful for understanding why firms had not adopted management practices
before.

III.D. Small sample size

The focus on large firms meant we had to work with a small sample of firms. This raises three
broad concerns. A first potential concern is whether the sample size is too small to identify
significant impacts of management. A second is what type of statistical inference is appropriate
given the sample size. Third, the sample may be too small to be representative of large firms in
developing countries. We discuss each concern in turn and the steps we have taken to try and
address some of these shortcomings.
16
Treatment and control plants were never in the same firms. The 6 control plants were randomly selected first, and the
14 treatment firms then randomly selected from the remaining firms which did not have a control plant.
17
Each wave had a one-day kick-off meeting with all the firms, involving presentations from a range of senior partners
from the consulting firm. This helped impress the firms with the expertise of the consulting firm and highlighted the
huge potential for improvements in management. This meeting involved a project outline, which was slightly different
for the treatment and control firms because of the different interventions. Since we did not tell firms about the existence
of treatment and control groups we could not mix the treatment and control groups.

10
Significance of results:
Even though we have only 20 experimental plants across 17 firms we obtain statistically significant
results. There are five reasons for this. First, these are large plants with about 80 looms and about
130 employees each, so that idiosyncratic shocks – like machine breakdowns or worker illness –
tend to average out.18 Second, the data were collected directly from the machine logs, so have very
little (if any) measurement error. Third, the firms are homogenous in terms of size, product, region
and technology, so that most external shocks can be controlled for with time dummies. Fourth, we
collected weekly data, which provides high-frequency observations over the course of the treatment.
Finally, the intervention was intensive, leading to large treatment effects – for example, the point
estimate for the reduction in quality defects was over 50%.

Statistical inference:
A second concern is over using standard statistical tests, which assume an asymptotically normal
distribution of errors across firms (the N dimension). We use three alternatives to address this
concern: First, we use firm-clustered bootstrap standard errors (see Bertrand, Duflo and
Mullainathan, 2004, and Cameron et al, 2008). Second, we implement permutation procedures (for
both the ITT and IV estimators) that have exact finite sample size and do not rely upon any
asymptotic approximations. Third, we exploit the fact we have a large T sample to implement
procedures that rely upon asymptotic approximations along the time dimension with a fixed N.

Permutation Tests: Permutation tests use the fact that order statistics are sufficient and complete
statistics to derive critical values for test statistics. We first implement this for the null hypothesis of
no treatment effect against the two sided alternative for the Intent to Treat (ITT) parameter. This
calculates the ITT coefficient for every possible combination of 11 treatment firms out of our 17
total firms (we run this as the firm rather than plant level to allow for firm-level correlations in
errors). Once this is calculated for the 12,376 possible treatment assignments (17 choose 11) the
2.5% and 97.5% confidence intervals are calculated as the 2.5th and 97.5th percentiles of the
treatment impact. A treatment effect outside these bounds can be said to be significant at the 5%
level.

Permutation tests for the instrumental variables estimator are more complex. For these we
implement an IV version of the permutation procedure based on the proposals in Greevy et al
(2004) and Andrews and Marmer (2008). The basic idea is to consider an outcome , endogenous
regressor , a randomized treatment assignment , and the model . Suppose that we
are interested in testing the null hypothesis that . Under the null hypothesis is
19
independent of the random treatment assignment variable . This suggests using measures of the
dependence between Z and as the basis of a testing procedure. The problem is now reduced
to testing for the independence between and Z, for which again we can use a permutation
test that has exact finite sample size (see Appendix B for more details).

T-asymptotic clustered standard errors: An alternative approach is to use asymptotic estimators that
exploit our large time dimension for each firm. To do this we use the recent results by Ibramigov

18
To illustrate this note that, for example, we actually have more employees in our 20 plants than in the 391 micro-
enterprises in De Mel et al. 2008.
19
If the null hypothesis is false then which will be correlated with
since the endogenous regressor is correlated with the instrument.

11
and Mueller (2009) (henceforth IM) to implement a t-statistic based estimator that is robust to
substantial heterogeneity across firms as well as to considerable autocorrelation across observations
within a firm. The IM approach requires estimating the parameter of interest separately for each
treatment firm and then treating the resultant set of 11 estimates as a draw from a t distribution with
10 degrees of freedom. IM show that such a procedure is valid in the sense of having correct size
(for fixed N) so long as the time dimension is large enough that the estimator for each firm can be
treated as a draw from a normal distribution. In our application we have on average over 100
observations for each firm, so this requirement is likely to be met.

Representativeness of the sample:

A third concern with our small sample is whether it is representativeness of large firms in
developing countries. In part this concern represents a general issue for field experiments, which are
often run on individuals, villages or firms in particular regions or industries. In our situation we
focus on one region and one industry, albeit India’s commercial hub (Mumbai) and it’s largest
industry (textiles). Comparing our sample to the population of large (100 to 5000 employee) firms
in India, both overall and in textiles, suggests that while our sample is small it is at least broadly
representative in terms of management practices (see Figure 1). In section V.D we also report
results on a plant by plant basis to further demonstrate these are not driven by any particular outlier
amongst the treatment or control plants. As such while we have a small sample the results are
relatively stable across the individual plants within the sample.

IV. THE IMPACT ON MANAGEMENT PRACTICES

In Figure 2 we plot the average management practice adoption of the 38 practices for the 14
treatment plants, the 6 control plants and the 8 non experimental plants. This data is shown at 2
month intervals before and after the diagnostic phase. Data from the diagnostic phase onwards was
compiled from direct observation at the factory. Data from before the diagnostic phase was
collected from detailed interviews of the plant management team based on any changes to
management practices during the prior year. Figure 2 shows five key results:

First, all plants started off with low baseline adoption rates of the set of 38 management practices. 20
Among the 28 individual plants the initial adoption rates varied from a low of 7.9% to a high of
55.3%, so that even the best managed plant in the group had just over half of the 38 key textile
manufacturing management practices in place. This is consistent with the results on poor general
management practices in Indian firms shown in Figure 1. For example, many of the plants did not
have any formalized system for recording or improving production quality, which meant that the
same quality defect could arise repeatedly. Most of the plants also had not organized their yarn
inventories, so that yarn stores were mixed by color and type, without labeling or computerized
entry. The production floor was often blocked by waste, tools and machinery, impeding the flow of
workers and materials around the factory.

20
The difference between the treatment, control and other plant groups is not statistically significant, with a p-value on
the difference of 0.248 (see Table A1).

12
Second, the intervention did succeed in changing management practices. The treatment plants
increased their use of the 38 management practices over the period by 37.8 percentage points on
average (an increase from 25.6% to 63.4% of practices implemented).

Third, the treatment plants’ adoption of management practices occurred gradually. In large part this
reflects the time taken for the consulting firm to gain the confidence of the firm’s directors. Initially
many directors were skeptical about the suggested management changes, and they often started by
piloting the easiest changes around quality and inventory. Once these started to generate substantial
improvements, these changes were rolled out and the firms started to consider introducing the more
complex improvements around operations and HR.

Fourth, the control plants, which were given only the 1 month diagnostic, increased their adoption
of these management practices, but by only 12% on average. This is substantially less than the
increase in adoption in the treatment wave firms, indicating that the four months of the
implementation the treatment plants received were important in changing management practices21.
The control firms typically did not adopt the more complex practices like daily quality meetings,
formalizing the yarn monitoring process or defining roles and responsibilities.

Fifth, the non experimental plants also saw a substantial increase in the adoption of management
practices. In these 8 plants the management adoption rates increased by 11%. Most of this increase
was driven by the 5 non experimental plants which were in the same firm as the treatment plants,
which increased the adoption of practices by 17.5%, compared with the 3 non experimental plants
which were in the same firm as the control plants which increasing their adoption by just 1%. This
spillover of management practices was driven by the directors copying the new management
practices from their experimental plants to their other plants.

IV.A. Management practice spillovers across plants within firms

To test formally whether the intervention differentially changed management practices between the
treatment and control plants we run the following plant-level panel regression:

MANAGEMENTi,t = αi + βt + λ1OWN_TREATi,t + λ2SPILLOVER_TREATi,t + εi,t (1)

where MANAGEMENTit is the fraction of the 38 practices adopted by plant I at date t, αi are plant
fixed effects, βt are calendar month fixed effects, OWN_TREATi,t = log(1+cumulative months since
the implementation phase began), and SPILLOVER_TREATi,t = log(1+cumulative months since
implementation began in all other plants in the same firm). We use this logarithmic functional form
because of concave adoption path of management practices shown in Figure 2. The parameter λ1
estimates the semi-elasticity of the plants management practices with respect to the months of their
own on-site consulting, while λ2 estimates the semi-elasticity of spillovers from on-site consulting
plants in other plants within the firm. The standard errors are bootstrap clustered by firm.

The results are shown in Table 2. We report in column (1) that management practices significantly
respond to own plant treatment, rising by about 0.109 for every unit change in log(1+months
treatment). We also see a response of 0.041 to log(1+months treatment) in other plants within the
same firm. This coefficient is just over one third of the magnitude of the direct impact, suggesting
21
This is consistent with Nelson and Winter’s (1982) discussions of tacit knowledge and transferable routines.

13
substantial spillovers of management practices across plants within the same firm. In column (2) we
add the three month lagged spillover term to investigate the timing of any potential spillover and
find the lagged term dominates. This is consistent with a delay in transferring management practices
across plants.22 This arises because the firms’ directors typically evaluate the impact of the new
management practices in their treatment plants before transferring these over to their other plants. In
column (3) we use just the three month lag and find a coefficient of 0.047, at almost half the direct
effect. In columns (4) and (5) we see that using an even longer six-month lags leads to similar
coefficient of 0.050. So whatever the exact specification, this data provides evidence of gradual
spillovers of better management practices across plants within firms. Importantly for our study,
these results also show that the experiment differentially changed management practices between
treatment and control plants, providing variation which we can use to examine the impacts of
management on performance.

V. THE IMPACT OF MANAGEMENT ON PERFORMANCE

The previous literature has shown a strong correlations between management practices and firm
performance in the cross-section, with a few papers (e.g. Ichniowski et al. 1998) also showing this
in the panel.23 Our unique panel data on management practices and plant level performance,
coupled with the experiment, enables us to confirm to what extent these results are causal.

We begin with a panel fixed-effects specification:

OUTCOMEi,t = αi + βt + θMANAGEMENTi,t+νi,t (2)

where outcome will be one of the three key performance metrics of quality, inventory and output.
The concern is that management practices are not exogenous to the outcomes that are being
assessed, even in changes. For example, a firm may only start monitoring quality when it is starting
to experience a larger than usual number of defects, which would bias the fixed-effect estimate
towards finding a negative effect of better management on quality. Or firms may start monitoring
quality as part of a major upgrade in worker quality and equipment, in which case we would
misattribute quality improvements from better capital and labor to better management.

To overcome this endogeneity problem, we instrument the management practice score with
log(1+weeks of treatment). The exclusion restriction is that the intervention affected the outcome of
interest only through its impact on management practices, and not through any other channel. A
justification for this assumption is that the consulting firm focused entirely on the 38 management
practices in their recommendations to firms, and firms did not buy new equipment or hire new labor
as a result of the intervention during the period of our study. The IV estimator will then allow us to
answer the headline question of this paper – does management matter?

22
In comparison the 3 month lagged own plant consulting term was actually negative and insignificant when added, so
that the cross-plant spillovers is delayed while the within plant treatment effect seems to happen instantly.
23
Note that most papers using repeated surveys have found no significant panel linkage between management practices
and performance (Cappelli and Neumark (2001) and Black and Lynch (2004)).

14
If the impact of management practices on plant-level outcomes is the same for all plants, then the
IV estimator will provide a consistent estimate of the marginal effect of improvements in
management practices, telling us how much management matters for the average firm participating
in the study. However, if the effects of better management are heterogeneous, then the IV estimator
will consistently estimate a local average treatment effect (LATE). The LATE will then give the
average treatment effect for plants which do change their management practices when offered free
consulting. If plants which stand to gain more from improving management are the ones who
change their management practices most as a result of the consulting, then the LATE will exceed
the average marginal return to management. It will understate the average return to management if
instead the plants that change management only when free consulting is provided are those with the
least to gain.

There was heterogeneity in the extent to which treatment plants changed their practices, with the
before-after change in average total management practice score ranging from 26.3 to 60 percentage
points. The feedback from the consulting firm was that to some extent it was firms with the most
unengaged, uncooperative managers who changed practices least, suggesting that the LATE may
underestimate the average impact of better management if these firms have the largest potential
gains from better management. Nonetheless, we believe the LATE to be a parameter of policy
interest, since if governments are to employ policies to try and improve management, information
on the returns to better management from those who actually change management practices when
help is offered is informative.

We can also directly estimate the impact of the consulting services intervention on management
practices via the following equation:

OUTCOMEi,t = ai + bt + cTREATi,t + ei,t (3)

Where TREATi,t is a 1/0 variable for whether plants have started the implementation phase or not.
The parameter c then gives the intention to treat effect (ITT), which is the average impact of the
intervention in the treated plants compared to the control plants.

In all cases we include plant and time fixed effects, and we bootstrap cluster the standard errors at
the firm level. We have daily data on many outcomes, but aggregate them to the weekly level to
reduce higher-frequency measurement errors.

V.A Quality
Our measure of quality is the Quality Defects Index (QDI), a weighted average score of quality
defects, which is available for all but one of the plants. Higher scores imply more defects. Figure 3
provides a plot of the QDI score for the treatment and control plants relative to the start of the
treatment period. This is September 2008 for Wave 1 treatment, April 2009 for Wave 2 treatment
and control plants.24 This is normalized to 100 for both groups of plants using pre-treatment data.
To generate point-wise confidence intervals we block bootstrapped over firms.

It is very clear the treatment plants started to reduce their QDI scores (i.e. improve quality)
significantly and rapidly from about week 5 onwards, which was the beginning of the

24
Since the control plants have no treatment period we set their timing to zero to coincide with the 10 Wave 2 treatment
plants. This maximizes the overlap of the data.

15
implementation phase following the initial 1 month diagnostic phase. The control firms also showed
a mild downward trend in their QDI scores from about week 30 onwards, consistent with their
slower take-up of these practices in the absence of a formal implementation phase. These
differences in trends between the treatment and control plants are also significant, as indicated by
the non-overlapping 95% confidence intervals towards the end of the period.

Table 3 in columns (1) to (3) examines whether management practices improve quality using a
regression analysis. In column (1) we present the fixed-effects OLS results which regress the
monthly log(Quality Defects Index) score on plant level management practices, plant fixed effects,
and a set of monthly time dummies. The standard errors are bootstrap clustered at the firm level to
allow for any potential correlation across different experimental plants within the same firm. The
coefficient of -0.561 implies that increasing the adoption of management practices by 10 percentage
points would be associated with a reduction of 5.61% in the quality defects index.

In Table 5, column (2), we instrument management practices using the experimental intervention to
identify the causal impact of better management on quality. After doing this we see a significant
point estimate of -2.028, suggesting that increasing the management practice adoption rate by 10%
would be associated with a reduction in quality defects of 20.3%. The large rise in the point
estimate from the OLS to the IV estimator suggests firms are endogenously adopting better
management practices when their quality starts to deteriorate. There was some anecdotal evidence
for the latter, in that the consulting firm reported plants with improving quality were often less keen
to implement the new management practices because they felt these were unnecessary. This
suggests that the fixed-effects estimates for management and performance in prior work like
Ichniowski, Prennushi and Shaw (1997) may be substantially underestimating the true impact of
management on performance.

The reason for this large effect is that measuring defects allows firms to address quality problems
rapidly. For example, a faulty loom that creates weaving errors would be picked up in the daily QDI
score and dealt with in the next day’s quality meeting. Without this, the problem would often persist
for several weeks since the checking and mending team had no system (or incentive) for resolving
defects. In the longer term the QDI also allows managers to identify the largest sources of quality
defects by type, design, yarn, loom and weaver, and start to address these systematically. For
example, designs with complex stitching that generate large numbers of quality defects can be
dropped from the sales catalogue. This ability to improve quality dramatically through systematic
data collection and evaluation is a key element of the successful lean manufacturing system of
production (see, for example, Womack, Jones and Roos, 1992).

Finally, in column (3) we look at the intention to treat (ITT), which is the average reduction in the
quality defects index in the period after the intervention in the treatment plants versus the control
plants. We see this is associated with a 32% (exp(-.386)-1) reduction in the QDI index.

At the foot of table 5 we also present our Ibragimov-Mueller (IM) and permutation alternative
significance tests. First, looking at the IM tests that exploit asymptotics in T rather than N , we find
that the IV and ITT results are both significant at the 5% level (zero is outside the 95% confidence
intervals). For the standard permutation tests the ITT is again significant at the 5% level (the p-
value is 0.0168), while for the IV-permutation tests the estimate is significant at the 10% level
although not at the 5% level. So overall the small sample statistical tests also find the IV and ITT
effects of management on quality significant at the 10% level and usually also the 5% level.

16
V.B Inventory
Figure 4 shows the plot of inventory levels over time for the treatment and control groups. It is clear
that after the intervention the inventory levels in the treatment group fall relative to the control
group, with this being point-wise significant by about 30 weeks after the intervention.

The reason for this effect is that these firms were carrying about 4 months of inventory on average
before the intervention, including a large amount of dead-stock. Often, because of poor records and
storage practices, firms did not even know they had these stocks. By cataloguing the yarn and
sending the shade-cards to the design team to include in new products25, selling dead yarn stock,
introducing restocking norms for future purchases, and monitoring inventory on a daily basis, the
firms dramatically reduced their inventories. But this took time as the reduction in inventories
primarily arose from lowering stocking norms and using old yarn for new products.

Table 5 columns (4) to (6) shows the regression results for raw material (yarn) inventory. In all
columns the dependent variable is the log of raw materials, so the coefficients can be interpreted as
the percentage reduction in yarn inventory. The results are presented for the 18 plants for which we
have yarn inventory data (two plants do not maintain yarn stocks on site). In column (4) we present
the fixed-effects results which regresses the monthly yarn on the plant level management practices,
plant fixed effects, and a set of monthly time dummies. The coefficient of -0.639 says that
increasing management practices adoption rates by 10 percentage points would be associated with a
yarn inventory reduction of about 6.39%. In Table 5, column (5s), we see the impact of
management instrumented with the intervention displays a point estimate of -0.929, again somewhat
higher than the FE estimates in column (1). Again, the IV estimator is higher than the OLS
estimator, suggesting that the adoption of better management practices may be endogenous, driven
in part by poor inventory performance. In column (6) we see the intervention is associated with an
average reduction in yarn inventory of (exp(-.179)-1=) 16.4%.

These numbers are substantial, but in fact US automotive firms achieved much greater reductions in
inventory levels (as well as quality improvements) by adopting Japanese lean manufacturing
technology beginning in the 1980s. Many firms reduced inventory levels from several months to a
few hours by moving to just-in-time production (Womack, Jones and Roos, 1991).

Finally, as with the quality defects estimates the IM confidence intervals for the IV estimator find
the coefficient significance at the 5% level. However, the IV permutation tests can not exclude zero
so are not finding significance at standard levels. Looking at the ITT coefficient we see that under
IM the results are significant at the 10% level although again not significant using the standard
permutation tests.

V.C Output
In Figure 5 we plot output over time for the treatment and control plants. Output is measured in
physical terms, as production picks. The results here are less striking, although output of the
treatment plants has clearly risen on average relative to the control firms, and this difference is
point-wise statistically significant in some weeks towards the end of the period.

25
Shade cards comprise a few inches of sample yarn, plus information on its color, thickness and material. These are
sent to the design teams in Mumbai who use these to design new products using the surplus yarn.

17
In columns (7) to (9) in table 5 we look at this in a regression setting with plant and time dummies.
In column (7) we see that for the OLS specification increasing the adoption of management
practices by 10 percentage points would be associated with a 1.27% increase in efficiency. In
column (8), we see the impact of management instrumented with the intervention displays a higher
significant point estimate of 0.346, suggesting a 10% increase in management adoption would lead
to a 3.46% increase in output. As with quality and inventory the IV estimator is again notably
higher than the OLS estimator, again indicating an endogenous adoption of better management
when output falls.

Finally, in column (9) we look at the intention to treat (ITT) and see a point estimate of 0.056,
implying a 5.4% increase in output (exp(0.056)-1), although this only significant at the 11% level.
Looking at the small-sample standard errors we find both the IM and permutation tests are,
however, significant at the 5% level for both the IV and ITT estimates.

There are several reasons for these increases in output. Undertaking routine maintenance of the
looms reduces breakdowns that stop production. Collecting and monitoring the breakdown data
frequently also helps highlight looms, shifts, designs and yarn-types that are associated with more
breakdowns thereby facilitating pro-active corrective measures. Visual displays around the factory
floor together with the incentive schemes based on these performance metrics motivate workers to
improve operating efficiency. Finally, keeping the factory floor clean and tidy reduces the number
of accidents, reducing incidents like tools falling into machines or fires damaging equipment. Again
the experience from lean manufacturing is that the collective impact of these procedures can lead to
extremely large improvements in operating efficiency, raising output levels.

V.D Results by plant

We can also examine the difference in quality, inventory and output after treatment on a plant by
plant basis. To do this Figures 6 plots the histograms of the before to after changes in our
performance measures for the treatment and control plants. It can be seen that no one outlier plant is
driving these differences, with all treatment plants improving their quality (top-left plot), nine of the
treatment plants improving their inventory (top-right plot) and all treatment plants improving their
output (bottom left plot). In comparison the control plants appear pretty randomly distributed
around the zero impact point. We can also test the statistical difference of these changes between
the two groups, and find the p-value on the difference in differences is 0.035 for quality, is 0.096 for
inventory and 0.010 for output.26

V.E Are the improvements in performance due to Hawthorne effects?

Hawthorne effects are named after a series experiments carried out at the Hawthorne Works in the
1920s and 1930s. The results apparently showed that simply running experiments led to an
improvement in performance, with the most cited result being that both reducing and increasing
ambient light levels led to higher productivity. While these putative Hawthorne effects in the
original experiments have long been disputed (e.g. Levitt and List, 2009), there remains a serious
concern that some form of the Hawthorne effect is causing our observed increase in plant
performance.

26
Formally, we test this by regressing the 20 plant level differences on a 1/0 dummy variable for being a treatment firm,
and report the p-value on that dummy, clustering at the parent firm level.

18
However, we think this is unlikely, for a series of reasons. First, our control plants also had the
consultants on site over a similar period of time as the treatment firms. Both sets of plants got the
initial diagnostic period and the follow-up measurement period, with the only difference being the
treatment plants also got an intensive intermediate 4 month implementation stage while the control
plants had briefer, but frequent, visits from the consultants collecting data. The control plants were
not told they were in the control group. Hence, it cannot be simply the presence of the consultants
or the measurement of performance that generated the improvement in performance. Second, the
improvements in performance took time to arise and they arose in quality, inventory and efficiency,
where the majority of the management changes took place. Third, these improvements persisted for
many months after the implementation period, so are not some temporary phenomena due to
increased attention. Finally, the firms themselves also believed these improvements arose from
better management practices, which was the motivation for them spreading these practices out to
their other plants not involved in the experiment.

VI. IMPACT OF MANAGEMENT ON ORGANIZATIONAL

STRUCTURE AND COMPUTERIZATION

V.A The impact of management practices on firm organization

Although our interventions were never intended to change the treatment firms’ organizational
design, theory gave us some reason to believe that organizational changes might follow from
increasing the amount and quality of information available to decision makers. In recent years a
large theoretical literature on the economics of organization has developed dealing with the locus of
decision-making within firms.27 However, this literature does not lead to clear-cut predictions about
the effects of increased availability of information to managers. On the one hand, models like
Garicano (2000) of hierarchy as a specialization in knowledge acquisition suggest that more
decisions ought to be taken at lower levels if the amount of information available to all levels is
increased. Similarly, a standard agency perspective might also suggest that more decisions would be
delegated if new or more accurate performance measures become available, especially if (as in our
sample) the directors are under significant time constraints. However, to the extent that the plant
managers were initially better informed than their bosses by virtue of being closer to the operations,
the availability of the better measures might have reduced their information advantage, favoring the
directors’ making more decisions. But while the theoretical literature is expansive, the empirical
literature is extremely limited.

To measure decentralization we collected data on eight variables: the locus of decision-making for
weaver hiring, manager hiring, spares purchases, maintenance planning, weaver bonuses,
investment, and departmental co-ordination, and the number of days per week the owner spends at
the factory. Because firms’ organizational designs change slowly over time, we collected this data
at lower frequencies – pre-intervention, in March 2010 and in August 2010. For every decision
except investment and days at the factory we scored decentralization on a 1 to 5 scale, where 1 was
defined as no authority of the plant manager over the decision and 5 as full authority (see Appendix
Table A2 for the full survey and Table A4 for descriptive statistics). These questions and scoring

27
For recent surveys see Garicano and Van Zandt (2011), Mookherjee (2011), and Bolton and Dewatripont (2011).

19
were based on the survey methodology in Bloom, Sadun and Van Reenen (2009), which measured
decentralization across countries and found developing countries like India, China and Brazil
typically have very centralized decision-making within firms. The measure of the decentralization
for investment was in terms of “The largest expenditure (in rupees) a plant manager (or other
managers) could typically make without a Directors signature”, which had an average of 12,608
rupees (about $250). Finally, the number of days the owners spend each week at the factory is a
revealed preference measure of decentralization. The owners are usually located either at their head-
offices in Mumbai (which they prefer as it reduces their commute) or at the factory (if it needs more
direct management).

To combine all eight decentralization measures into one index we took the principal factor
component, which we called the decentralization index. Changes in this index were strongly and
significantly correlated with changes in management across firms. Firms which had substantial
improvements in management practices during the experiment also tended to delegate more
production decisions to their plant managers. Table 4 looks at this in a regression format:

DECENTRALIZATIONi,t = ai + bt +cMANAGEMENTi,t + ei,t (3)

where DECENTRALIZATION is our index of plant decentralization, and ai and bt are plant fixed
effects and time dummies. In column (1) we run the OLS estimation and find a significant and
positive coefficient, indicating that firms which improved their management practices during the
experiment have also delegated more decisions to their plant managers. Given that the
decentralization index has a standard deviation of 1 the magnitude of this coefficient is large –
increasing the adoption of management practices by 37.8% (the mean change for the treatment
group) is associated with a 0.55 standard-deviation change increase in decentralization. In column
(2) we run the IV estimation, using the log(1+weeks treatment) as the instrument, and again find a
positive and significant impact. Finally, in column (3) we report a positive ITT.

Our evidence from discussion with the owners is that better management leads to decentralization
because it increases their trust in their plant managers. In large part this is because the improved
monitoring of the factory operations allows them to delegate more decision making without fear of
being exploited (the monitoring channel in our principal-agent group of organizational theories).
For example, with daily inventory, quality and output data it is harder for the factory manager to
appropriate inventory or output without detection. A second channel seems to be that better
management practices enables the plant managers to more effectively run the factory without
assistance from the owner. For example, after introducing daily factory meetings the owners told-us
they felt less obliged to visit the factories daily to oversee production (the informational channel in
our first group of theories).

It is worth noting that even these decentralizing Indian factories are still extremely centralized
compared to factories in Europe and the US. For example, using the Bloom, Sadun and Van Reenen
(2009) data we know that plant managers in developed countries are typically able to invest about
$52,000 without central clearance, compared to about $250 in our India firms.

VI.B The impact of management practices on computerization

One of the major topics over the last decade has been the relationship between IT and productivity.
Until the 1990s, evidence on the impact of computers on productivity was so hard to find that
Robert Solow famously quipped in 1987 that “you see computers everywhere but in the

20
productivity statistics”. More recently, however, the paradox has reversed, with a growing literature
now finding that the productivity impact of IT is substantially larger than its cost share (e.g.
Bresnahan, Brynjolfsson and Hitt, 2002, and Brynjolfsson and Hitt, 2003). The literature argues this
is because IT is complementary with modern management and organizational practices, so that as
firms invest in IT they also improve their management practices. This leads to a positive bias on IT
in productivity estimates as management and organizational practices are typically an unmeasured
residual.28 But none of this literature has any experimental evidence.

A second related IT literature has argued that skill biased technical change (SBTC) has been the
major factor driving the increase in income inequality observed in the US and most other countries
since the 1970s (see surveys in Acemoglu 2002 and Autor, Katz and Kearney 2008). But SBTC is
usually inferred as the residual in inequality regressions, with rather limited direct evidence on
specific skill-biased technologies. Our experimental changes in management practices are skilled-
biased, in that computer users in India are relatively skilled due to the need for literacy and
numeracy. As a result modern management practices are a skill-biased technology, driving both the
use of computers and the demand for skilled workers.

So to investigate the potential complementarity between IT and management practices we collected

computerization data on nine aspects of the plants, covering the use of Electronic Resource
Planning (ERP) systems, the number of computers, the age of the computers, the number of
computer users, the total hours of computer use, the connection of the plant to the internet, the use
of e-mail by the plant manager and the director, the existence of a firm website and the depth of
computerization of production decisions (see Appendix Table A3 for the full survey and Table A4
for descriptive statistics). As with the organizational changes we collected this data once from
before the intervention, in March 2010 and in August 2010. Even in table A4 it is readily apparent
that as firms adopted more modern management practices they significantly increased the
computerization of their operations. Table 4 looks at this in a regression format:

COMPUTERIZATIONi,t = ai + bt +cMANAGEMENTi,t + ei,t (4)

where COMPUTERIZATION is measured in terms of the number of computer users (in columns
(4) to (6)) or in terms of the overall computerization index (in columns (7) to (9)).

In column (4) we see that the full adoption of all management practices is associated with an
increase of 16.76 hours of computer use a week, an average rise of over 100% given the pre-sample
mean was 13.66 hours per week. In columns (5) and (6) we report the IV and ITT estimates, which
show a similar result. In columns (7) to (9) we report similar OLS, IV and ITT results for the
computerization index, which is a broader measure of computer use, and again see highly
significant increases from the management intervention.

For context we should note, however, that even after the intervention these firms had extremely low
rates of computer use compared to firms in developed countries. For example, the median numbers
of computers per employee in our Indian plant was 0.023 compared to 1.18 in similar sized US and
European factories (see Bloom, Sadun and Van Reenen (2009)). And in US and European factories
all plant managers have a personal e-mail, compared to just 25% in India. Our results suggest one

28
See, for example, Bartel, Ichniowski and Shaw (2007) and Bloom, Sadun and van Reenen (2009).

21
reason for this low overall level of computer use in developing countries may be the lack of modern
management practices which these technologies support.

VII. WHY DO BADLY MANAGED FIRMS EXIST?

Given the evidence in section (IV) on the substantial impact of better management practices on
plants’ quality, inventory and output, the obvious question is whether these management changes
increased profitability, and if so why where these not introduced before.

VII.A. The estimated impact of management practices on profits and productivity

Profits:
Overall we estimate a total increase in profits of around $228,000, with our calculations outlined in
Table A5. Firms did not provide us with any profit and loss accounts, so we estimated this from the
quality, inventory and efficiency improvements.29 Our methodology is very simple: for example, if
a given improvement in practices is estimated to reduce inventory stock by X tons of yarn, we map
this into profits using conservative estimates of the cost of carrying X tons of yarn. Or if it reduces
the numbers of hours required to mend defects we estimated this reduction in hours on the firms
total wage bill. These estimates are medium-run because, for example, it will take a few months for
the firms to reduce their mending manpower.

These estimates for increases in profits are potentially biased. They are a downward biased because
we take firms’ choice of capital, labor and product range as given. But in the long-run the firms can
re-optimize, for example, with more machines per weaver if quality improves (as dealing with
breakdowns is time consuming). Furthermore, many of the management practices are
complementary, so they are much more effective when introduced jointly (e.g. Milgrom and
Roberts, 1990). However, the intervention time-horizon was too short to change many of the
complementary human-resource practices, so the full rewards would not be realized. The
intervention was also narrow in focus in that other management practices around activities like
finance, strategy, marketing and procurement were not addressed. They are clearly upward biased if
once the consultants leave the factory the firms backslide on the management changes.

To estimate the net increase in profit for these improvements in management practices we also need
to calculate the costs of these changes (ignoring for now any costs of consulting). These costs were
extremely small, averaging less than $3000 per firm.30 So given the $250,000 this consulting would
have cost these firms to buy, this implies returns of about 90% on such an investment, at least in the
first year.

Productivity:

29
We could obtain the public profit and loss accounts, but it was unclear how accurate these were, and they were not at
the plant level. We did not ask firms for their private profit and loss accounts (if they even kept them) as they would
have been likely to refuse, given their fears over the information leaking out to the Indian tax authorities.
30
The $35 of extra labor to help organize the stock rooms and clear the factory floor, about $200 on plastic display
boards, about $200 for extra racking for stores rooms, about $1000 on rewards, and about $1000 for extra computer
equipment (this is bought second hand).

22
We estimate a total increase in productivity of 11.1%, detailed in Table A5. Our methodology is
very simple, assuming a constant-returns-to-scale Cobb-Douglas production function:

Y=ALαK1-α (1)

where Y is value-added (output – materials and energy costs), L is hours of work and K is the net
capital stock. Using equation (1) we can back out changes in productivity after estimating changes
in output and inputs. So, for example, reducing the yarn inventory by 16.4% lowers capital by 1.3%
(yarn is 8% of the total capital stock), increasing productivity by 0.6% (capital has a factor share of
0.42).

Our estimated productivity impact will be subject to a number of biases similar to those discussed
above for profitability. However, the relatively substantial 11% impact from this narrow short-run
management experiment suggests that bad management does play a potentially important role in
explaining the aggregate productivity gap between India and the US.

VII.B. Why are firms badly managed?

Given the evidence in section (VI.A) above on the large increase in profitability from the
introduction of these modern management practices, the obvious question is why had firms not
already adopted them? To investigate this we asked our consultants to document every other month
the reason for the non-adoption of any of the 38 practices in each plant. To do this consistently we
developed a flow-chart (see Exhibit 7) which runs through a series of questions to understand the
root cause for the non-adoption of each practice. They collected this data from extensive discussions
with owners, managers and workers, plus their own observations.

As an example of how this flow chart works, imagine a plant that does not record quality defects.
The consultant would first ask if there was some external constraint, like labor regulations,
preventing this, which we found never to be the case.31 They would then ask if the plant was aware
of this practice, which in the example of quality recording systems typically was the case, as it is a
well known practice. The consultants would then check if the plant could adopt the practice with the
current staff and equipment, which again for quality recording systems was always true. Then they
would ask if the owner believed it would be profitable to record quality defects, which was often the
constraint on adopting this practice. The owner frequently argued that quality was so good they did
not need to record quality defects. This view was mistaken, however, because while these plants’
quality might have been good compared to other low-quality Indian textile plants, it was very poor
by international standards. So, as shown in Figure 3, when they did adopt basic quality control
practices they substantially improved their production quality. So, in this case the reason for non-
adoption would be “incorrect information” as we believed the owner had incorrect information on
the cost-benefit calculation.

The overall results for non-adoption of management practices are tabulated in Table 5, for the
treatment plants, control plants and the non-experimental plants. This is tabulated at two-month
intervals starting the month before the intervention phase. The rows report the different reasons for
non-adoption as a percentage of all practices. From the table several results are apparent.

31
This does not mean labor regulations do not matter for some practices – for example firing underperforming
employees – but they did not directly impinge adopt the immediate adoption of the 38 practices.

23
First, a major initial barrier to the adoption of these modern management practices was a lack of
information about their existence. About 15% of practices were not adopted because the firms were
simply not aware of them. These practices tended to be the more advanced practices of regular
quality, efficiency and inventory review meetings, posting standard-operating procedures and visual
aids around the factory. Many of these are derived from the Japanese inspired lean manufacturing
revolution, and are common across Europe, Japan and the US but apparently have yet to permeate
Indian manufacturing.

Second, another major initial barrier was incorrect information, in that firms may have heard of
these practices but thought they did not apply profitably to them. For example, many of the firms
were aware of preventive maintenance but few of them thought it was worth doing this. They
preferred to keep their machines in operation until they broke down, and then repair them. This
accounted for slightly over 45% of the initial non-adoption of practices.

Third, as the intervention progressed the lack of information constraint was rapidly addressed.
However, the incorrect information constraints were harder to address. This was because the owners
had prior beliefs about the efficacy of a practice and it took time to change these. This was often
done using pilot changes on a few machines in the plant or with evidence from other plants in the
experiment. For example, the consultants typically started by persuading the managers to undertake
preventive maintenance on a set of trial machines, and once it was proven successful it was rolled
out to the rest of the factory. And as the consultants demonstrated the positive impact of these initial
practice changes, the owners increasingly trusted them and would adopt more of the
recommendations, like performance incentives for managers.32

Fourth, once the informational constraints were addressed, other constraints arose. For example,
even if the owners became convinced of the need to adopt a practice, they would often take several
months to adopt it. This was particularly pertinent in the non-experimental plants, where the
consultants were not on-site to drive the changes. A major reason is the owners were severely time
constrained, working an average of 68 hours per week already. There was also evidence of
procrastination in that some owners would defer on taking quick decisions. This matches up with
the evidence on procrastination in other contexts, for example African farmers investing in fertilizer
(Duflo, Kremer and Robinson, 2009) or growing pineapples (Conley and Udry, 2010).

Finally, somewhat surprisingly, we did not find evidence for the direct impact of a set of other
factors that have been highlighted in the literature on capital investment. One such factor is capital
constraints, which are a significant obstacle to the expansion of micro-enterprises (e.g. De Mel et al,
2008). Our evidence suggested that the relatively large firms that were involved in our experiment
were not cash-constrained. We collected data on all the investments for our 17 firms over the period
April 2008 until April 2010 and found the firms invested a mean (median) of $880,000 ($140,000).
For example, several of the firms were setting up new factories or adding machines, apparently
often financed by bank loans. Certainly, this scale of investment suggests that investment on the
scale of $2000 (the first-year costs of these management changes, ignoring the consultants’ fees) to

32
These sticky priors highlight one reason why management practices appear to take several years to change in the US
and Europe. The evidence on this is anecdotal, but for example, the private equity industry has a 3 year minimum
estimate for the time needed for a management turnaround. Similarly, consulting firms typically take at least 18 months
to execute large change management programs at their clients.

24
improve the factories’ management practices is unlikely to be directly impeded by financial
constraints.

Of course financial constraints could impede hiring in international consultants. The market cost of
our free consulting would be at least $250,000, and as an intangible investment it would be difficult
to collateralize.33 Hence, while financial constraints do not appear to directly block the implantation
of better management practices, they may hinder firms’ ability to improve their management using
external consultants. On the other hand, our estimates of the 90% return on this consulting to
improve management practices suggest cost recovery within eighteen months.

VII.C. How do badly managed firms survive?

We have shown that management matters, with improvements in management practices improving
plant-level outcomes. One response from economists might then be to argue that poor management
can at most be a short-run problem, since in the long run better managed firms should take over the
market. Yet many of our firms have been in business for 20 years and more.

One reason why better run firms do not dominate the market is constraints on growth through
managerial span of control. In every firm in our sample, only members of the owning family have
managerial positions with major decision-making power over finance, purchasing, operations or
employment. Non-family members are given junior managerial positions with power only over
basic day-to-day activities. The reason is the family members do not trust non-family members. For
example, they are concerned if they let their plant managers procure yarn they may do so at inflated
rates from friends and receive kick-backs.

A key reason for this inability to decentralize is the poor rule of law in India. Even if directors
found managers stealing, their ability to successfully prosecute them and recover the assets is
minimal because of the inefficiency of Indian courts. In contrast, in the US if a manager was found
stealing from a firm it is likely they could be successfully prosecuted and much of the assets
recovered. A compounding reason for the inability to decentralize in Indian firms is bad
management, as this means the owners cannot keep good track of materials and finance, so may not
even able to identify theft within their firms.34

As a result of this inability to delegate, firms can expand beyond the size that can be managed by a
single director only if other family members are available to share directorial duties. Thus, an
important correlate of firm size in our firms was the number of male family members of the owners.
For example, the number of brothers and sons of the leading director has a correlation of 0.689 with
the total employment of the firm, compared to a correlation of 0.223 with their average
management score across plants. In fact the best managed firm in our sample had only one (large)

33
Our international consulting firm estimated that to offer a standard consulting team to these firms at market rates
would cost at least $500,000. This is much more expensive than our costs per firm because: (I) we achieved substantial
scale economies from working with a large number of firms simultaneously; and (II) we had 50% rates on the
consultants and no partner charges.
34
Another compounding factor is these firms had poor human resources management practices. None of the firms had a
formalized development or training plan for their managers, and managers could not be promoted because only family
members could become directors. As a result managers lacked career motivation within the firm. In contrast, in Indian
software and finance firms place a huge emphasis on development and training to motivate employees, which is
essential for delegation without a strong legal system (see also Banerjee and Duflo (2000)).

25
production plant, in large part because the owner had no brothers or sons to help run a larger
organization. This matches the ideas of the Lucas (1978) span of control model, that there are
diminishing returns to how much additional productivity better management technology can
generate from a single manager. In the Lucas model, the limits to firm growth restrict the ability of
highly productive firms to drive lower productivity ones from the market. In our Indian firms, this
span of control restriction is definitely binding, so unproductive firms are able to survive because
more productive firms cannot expand.

Entry of new firms into the industry also appears limited by several factors, several related to the
difficulty of separating ownership from control. The supply of new firms is constrained by the
number of families with finance and male family members available to build and run textile plants.
Since other industries in India – like software and real-estate – are growing rapidly the
attractiveness of new investment in textile manufacturing is relatively limited. (Even our firms were
often taking cash from their textile businesses to invest in other businesses, like real-estate). Finally,
even if an entrant had funding, given the informational problems identified earlier, there is no
obvious guarantee its management practices would be better than the incumbent firms.

Finally, a 50% tariff on fabric imports insulates Indian textile firms against foreign competition.
Hence, the equilibrium appears to be that, with Indian wage rates being extremely low, firms can
survive with poor management practices. Because spans of control are constrained productive
incumbent firms are limited from expanding and so do not drive out the badly run firms. And
because entry is limited new firms do not enter rapidly. As a result the situation approximates a
Melitz (2003) style model where firms have very high decreasing returns to scale, entry rates are
low, and initial productivity draws are low (because good management practices are not
widespread). The resultant equilibrium has a low average level of productivity, a low wage level, a
low average firm-size, and a large dispersion of firm-level productivities.

VII.D. Why do firms not use more management consulting?

Finally, why do these firms not hire consultants themselves, given the large gains from better
management? The primary reason is these firms are not aware they are badly managed, as
illustrated in Table 5. Of course consulting firms could still approach firms for business, pointing
out that their practices were bad and offering to fix them. But Indian firms, much like US firms, are
bombarded with solicitations from businesses offering to save them money on everything from
telephone bills to raw materials, and so are unlikely to be receptive. Of course consulting firms
could offer to provide free advice in return for an ex post profit-sharing deal. But monitoring this
would be hard. First, how are the returns attributable to the consulting to be determined? In the
Indian context, firms very often conceal profits from the tax authorities and would be tempted to do
the same with partnering consultants. Moreover, the firm in such an arrangement might worry that
the consultant would twist its efforts to increase short-term profits at the expense of long-term
profits. Finally, many Indian firms are breaching tax, labor and health-and-safety laws and so are
reluctant to let unknown outsiders into their firms. Our project benefited from the endorsement of
Stanford and the World Bank, but a local firm offering free consulting would probably find it much
harder to gain the trust of potential clients.

26
 VIII. CONCLUSIONS

Management does matter. We implemented a randomized experiment that provided managerial

consulting services to textile plants in India. This experiment led to improvements in basic
management practices, with plants adopting lean manufacturing techniques that have been standard
for decades in the developed world. These improvements in management practice led to
improvements in product quality, reductions in inventory and increased efficiency, raising
profitability and productivity. Firms also delegated more decisions because the improved
informational flow from adopting modern management practices enabled the owners to reduce their
oversight of plant operations. At the same time computer use increased, driven by the need to
collect, process and disseminate data as required by modern management practices.

What are the implications of this for public policy? First, our results suggest that firms were not
implementing best practices on their own because of lack of information and knowledge, and that to
really improve management quality, firms needed persuasion that these practices would improve
productivity and how they could implement them. This suggests a need for better benchmarking
programs – to convince firms of the need to improve management practices – and knowledge and
training programs in India and other developing countries more generally. This would include high
quality educational programs to teach better management practices, and a more vibrant local
consulting industry with the ability to signal quality through reputation building. While both these
are private sector activities, they depend on the government for a regulatory environment which
makes entry easy and which allows quality to be the main determinant of success.

A second method for knowledge transference comes from the presence of multinationals. Indeed,
many of the consultants working for the international consulting firm hired by our project had
worked for multinationals in India, learning from their state-of-the-art manufacturing management
processes. Yet a variety of legal, institutional, and infrastructure barriers have limited the extent of
multinational expansion within India, limiting the spread of knowledge on better manufacturing
among the Indian managerial labor force. Abolishing tariffs would also help substantially, as Indian
firms would be driven to improve management practices to survive against lower cost imports from
countries like China and Vietnam.

Finally, our results also suggest that a weak legal environment has limited the scope for well-
managed firms to grow. So that improving the legal environment should encourage productivity-
enhancing reallocation, helping to drive out badly managed firms.

27
 BIBLIOGRAPHY
Acemoglu, Daron (2002) “Technical Change, Inequality and the Labor Market”, Journal of Economic
Literature, 40(1): 7-72.
Acemoglu Daron, Philippe Aghion, Claire Lelarge, John Van Reenen, and Fabrizio Zilibotti (2007)
“Technology, Information and the Decentralization of the Firm”, Quarterly Journal of Economics, 122(4),
1759–1799.
Aghion, Philippe, and Jean Tirole (1997) “Formal and Real Authority in Organizations”, Journal of Political
Economy, 105(1), 1-29.
Alonso, Ricardo, Wouter Dessein and Niko Matouschek (2008) “When Does Coordination Require
Centralization”, American Economic Review, 98(1), 145-179.
Andrews, D.W.K and Marmer, V. (2008) “Exactly Distribution-free Inference in Instrumental Variables
Regression with Possibly Weak Instruments, ” Journal of Econometrics. 142(1): 183-200
Autor, David, Katz, Lawrence, and Kearney, Melissa (2008), “Trends in US wage inequality: revising the
revisionists”, Review of Economics and Statistics, 90 (2): 300-323.
Baker, George, Robert Gibbons, and Kevin Murphy (1999) “Informal Authority in Organizations”, Journal
of Law, Economics, and Organization, 15(1), 56-73.
Baker, George, and Thomas Hubbard (2004) “Contractibility and Asset Ownership: On Board Computers
and Governance in US Trucking”, Quarterly Journal of Economics, 119(4), 1443-1479.
Banerjee, Abhijit and Ester Duflo (2000) “Reputation effects and the limits of contracting: a study of the
Indian software industry”, Quarterly Journal of Economics vol 115(3), pp. 989-1017.
Banerjee, Abhijit and Ester Duflo (2005) “Growth Through the Lens of Development Economics”, in
Philippe Aghion and Stephen Durlauf (eds), Handbook of Economic Growth, Volume 1 of Handbook of
Economic Growth, Chapter 7, pp. 473-552. Amsterdam: Elsevier.
Bartel, Ann, Casey Ichniowski and Kathryn Shaw, 2007. ‘How Does Information Technology Really Affect
Productivity? Plant-Level Comparisons of Product Innovation, Process Improvement and Worker Skills’,
Quarterly Journal of Economics, 122(4), 1721-1758.
Bertrand, Marianne and Antoinette Schoar (2003) “Managing with Style: The Effects of Managers on
Corporate Policy”, Quarterly Journal of Economics 118(4): 1169-1208.
Bertrand, Marianne and Duflo, Esther and Mullainathan, S. (2004) “How Much Should We Trust Difference-
in-Difference Estimators,” Quarterly Journal of Economics. 119(1): 249-275
Bester, C.A and Conley, T.G and Hansen, C.B. (2008) “Inference with Dependent Data Using Cluster
Covariance Estimators” Working Paper
Black, Sandra, and Lisa Lynch. 2001. “How to Compete: The Impact of Workplace Practices and
Information Technology on Productivity.” Review of Economics and Statistics, 88(3): 434-45.
Black, Sandra and Lisa Lynch. 2004. ‘What's Driving the New Economy? The Benefits of Workplace
Innovation’, Economic Journal, 114(493), 97-116.
Bloom, Nicholas, Raffaella Sadun, and John Van Reenen (2009) “Americans do IT Better: American
Multinationals and the Productivity Miracle”, forthcoming American Economic Review.
Bloom, Nicholas, Sadun, Raffaella, and John Van Reenen (2009) “The organization of firms across
countries”, NBER Working Paper No. 15129.
Bloom, Nicholas, and John Van Reenen (2007) “Measuring and Explaining Management Practices across
Firms and Countries”, Quarterly Journal of Economics, 122(4), 1341-1408.
Bloom, Nicholas, and John Van Reenen (2010) “Human Resource Management and Productivity”, draft
chapter for the Handbook of Labor Economics.
Bloom, Nicholas, and John Van Reenen (2010) “Why do management practices differ across firms and
countries”, Journal of Economic Perspectives, 24(1), 203-224.
Bresnahan, Timothy, Erik Brynjolfsson, and Lorin Hitt (2002) “Information Technology, Workplace
Organization and the Demand for Skilled Labor: Firm-level Evidence”, Quarterly Journal of Economics,
117(1), 339-376.

28
Bruhn, Miriam, Karlan, Dean and Schoar, Antoinette (2010), “The impact of offering consulting services to
small and medium enterprises: evidence from a randomized trial in Mexico”, mimeo.
Cappelli, P and Neumark, D. 2001. “Do ‘High Performance” Work Practices Improve Establishment-Level
Outcomes?” Industrial and Labor Relations Review, 737-775.
Clark, Greg (1987), “Why isn’t the whole world developed”, Journal of economic history, 47, pp. 141-173.
Conley, Tim and Udry, Christopher, (2010), “Learning about a new technology: pineapple in Ghana”,
American Economic Review 100(1): 35-69.
Davis, Steve, Haltiwanger, John and Schuh, Scott, (1996), Job Creation and Destruction, MIT Press.
De Mel, Suresh, David McKenzie and Christopher Woodruff (2008) “Returns to Capital in Microenterprises:
Evidence from a Field Experiment”, Quarterly Journal of Economics 113(4): 1329-1372.
Delery, John and Doty, Harold (1996), “Modes of theorizing in strategic human resource management: test
of universalistic, contingency and configurational performance predictions”, Academy of Management
Review, vol. 39(4), pp 802-835.
Drexler, Alejandro, Fischer, Greg and Schoar, Antoinette (2010), “Financial literacy training and rule of
thumbs: evidence from a field experiment”, mimeo.
Duflo, Esther, Kremer, Michael and Robinson, Jonathan (2009), “Nudging farmers to use fertilizer: theory
and experimental evidence from Kenya”, Harvard mimeo.
Foster, Lucia, John Haltiwanger and Chad Syverson (2008) “Reallocation, Firm Turnover, and Efficiency:
Selection on Productivity or Profitability?” American Economic Review, 98(1), 394-425
Garicano, Luis (2000) “Hierarchies and the Organization of Knowledge in Production”, Journal of Political
Economy, 108(5), 874-904.
Garicano, Luis and Van Zandt, Timothy (2010), “Hierarchy”, draft chapter for publication in Robert Gibbons
and John Roberts, editors, Handbook of Organizational Economics, Princeton University Press.
Gibbons, Robert and John Roberts (2010) The Handbook of the Organizational Economics, Princeton:
Princeton University Press.
Greevy, R. and Silber, J.H. and Cnaan A. and Rosenbaum, P.R. (2004) Randomization Inference with
Imperfect Compliance in the ACE-inhibitor after anthracycline randomized trial. Journal of the American
Statistical Association. 99(465): 7-15
Guadalupe, Maria and Julie Wulf (2007) “The Flattening Firms and Product Market Competition: The
Effects of Trade Costs and Liberalization”, Columbia University mimeo.
Hsieh, Chiang-Tai and Pete Klenow (2009), “Misallocation and Manufacturing TFP in China and India”,
Quarterly Journal of Economics 124(4): 1403-1448.
Hsieh, Chiang-Tai and Pete Klenow (2010), “Development accounting”, American Economic Journal:
Macroeconomics 2(1):207-223.
Huselid, Mark and Brian Becker, 1996. ‘Methodological Issues in Cross-sectional and Panel Estimates of the
Human Resource-firm Performance link’, Industrial Relations, 35, 400-422.
Ibragimov, R. and Muller, U.K. (2009) “t-statistic Based Correlation and Heterogeneity Robust Inference,”
Journal of Business and Economic Statistics.
Ichniowski, Casey, Kathryn Shaw and Giovanna Prenushi. (1997), “The Effects of Human Resource
Management: A Study of Steel Finishing Lines”, American Economic Review, 87(3), 291-313.
Jorgenson, Dale, Mun Ho and Kevin Stiroh, 2008. ‘A Retrospective Look at the US Productivity Growth
Resurgence’, Journal of Economic Perspectives, 22(1), 3-24.
Karlan, Dean and Martin Valdivia (2009) “Teaching Entrepreneurship: Impact Of Business Training On
Microfinance Clients and Institutions”, Review of Economics and Statistics, forthcoming.
Lazear, Edward, and Paul Oyer. 2009. “Personnel Economics.” in Robert Gibbons and John Roberts, eds.
forthcoming in Handbook of Organizational Economics.
Levitt, Steven and List, John, (2009), “Was there really a Hawthorne Effect at the Hawthorne Works? An
Analysis of the Original Illumination Experiments”, NBER WP15016.
Lucas, Robert E. (1978): “On the size distribution of business firms.” Bell Journal of Economics, 9:508-523.
MacDuffie, John Paul. (1995), “Human Resource Bundles and Management Performance: Organizational
Logic and Flexible Production Systesm in the World Auto Industry,” Industrial and Labor Relations
Review, 48(2): 197-221.

29
McKenzie, David (2009) “Impact Assessments in Finance and Private Sector Development: What have we
learned and what should we learn?”, World Bank Research Observer, forthcoming.
McKinsey Global Institute (2001) India: The Growth Imperative.
http://www.mckinsey.com/mgi/publications/India.asp
Milgrom, Paul and Roberts, John (1990), “The economics of modern manufacturing: technology, strategy
and organization”, American Economic Review, 80 (3), pp. 511-528.
Mookherjee, Dilip (2010) draft chapter for publication in Robert Gibbons and John Roberts, editors,
Handbook of Organizational Economics, Princeton University Press.
Nelson, Richard and Winter, Sydney (1982) An Evolutioinary Theory of Economics Change, Harvard
University Press.
Osterman, Paul, 1994. ‘How Common Is Workplace Transformation and Who Adopts It?’, Industrial and
Labor Relations Review, 47(2), 173-188 .
Pack, Howard (1987) Productivity, Technology, and Industrial Development: A Case Study in Textiles,
World Bank Publications.
Rajan, Raghuram, and Luigi Zingales (2001) “The Firm as a Dedicated Hierarchy: A Theory of the Origin
and Growth of Firms”, Quarterly Journal of Economics, 116(3), 805-851.
Rajan, Raghuram, and Julie Wulf (2006) “The Flattening Firm: Evidence from Panel Data on the Changing
Nature of Corporate Hierarchies”, Review of Economics and Statistics, 88(4), 759-773.
Syverson, Chad (2004a), “Market Structure and Productivity: A concrete example”, Journal of Political
Economy, 112(6), 1181-1222.
Syverson, Chad (2004b), “Product substitutability and Productivity Dispersion”, Review of Economics and
Statistics, 86 (2), 534-50.
Syverson, Chad (2010), ‘’What determines productivity at the micro level?”, draft manuscript for the Journal
of Economic Literature.
Taylor, Fredrick (1911), Principles of Scientific management, Harper and brothers, New York and London.
Walker, Francis (1887), “The source of business profits”, Quarterly Journal of Economics, 1(3), pp. 265-
288.
Womack, James, Jones, Daniel and Roos, Daniel (1991), “The machine that changed the world”, Harper
Collins publishers, New York, USA.
Wooldridge, J. (2004) “Cluster Sample Methods in Applied Econometrics,” American Economic Review.
Woodward J. 1958 Management and Technology, Cambridge: Cambridge University Press.

30
APPENDIX A: DATA
Our estimates for profits and productivity impacts are laid out in Table A5, with the methodology
outlined below. We calculate the numbers for the median firm.

A. Estimations of profitability and productivity impacts.

We first generate the estimated impacts on quality, inventory and efficiency. To do this we take the
Intention to Treat (ITT) numbers from Table 3, which shows a reduction of quality defects of 32% (exp(-
0.386)-1), a reduction in inventory of 16.4% (exp(-0.179)-1) and an increase in output of 5.4%
(exp(0.056)-1).

Mending wage bill:

Estimated by recording the total mending hours, which is 71,700 per year on average, times the mending
wage bill which is 36 rupees (about $0.72) per hour. Since mending is undertaken on a piece-wise basis –
so defects are repaired individually – a reduction the severity weighted defects should lead to a
proportionate reduction in required mending hours.

Fabric revenue loss from non grade-A fabric:

Waste fabric estimated at 5% in the baseline, arising from cutting our defect areas and destroying and/or
selling at a discount fabric with unfixable defects. Assume increase in quality leads to a proportionate
reduction in waste fabric, and calculate for the median firm with sales of $6m per year.

Inventory carrying costs:

Total carrying costs of 22% calculated as interest charges of 15% (average prime lending rate of 12%
over 2008-2010 plus 3% as firm-size lending premium – see for example
http://www.sme.icicibank.com/Business_WCF.aspx?pid), 3% storage costs (rent, electricity, manpower
and insurance) and 4% costs for physical depreciation and obsolescence (yarn rots over time and fashions
change).

Increased profits from higher output

Increasing output is assumed to lead to an equi-proportionate increase in sales because these firms are
small in their output markets, but would also increase variable costs of energy and raw-materials since the
machines would be running. The average ratio of (energy + raw materials costs)/sales is 63%, so the
profit margin on increased efficiency is 37%.

Labor and capital factor shares:

Labor factor share of 0.58 calculated as total labor costs over total value added using the “wearing
apparel” industry in the most recent (2004-05) year of the Indian Annual Survey of industry. Capital
factor share defined as 1-labor factor share, based on an assumed constant returns to scale production
function and perfectly competitive output markets.
APPENDIX B: ECONOMETRICS
We briefly outline in this section the v. The proposed procedure by Ibragimov-Mueller (2009) (IM) is
useful for our case where the number of entities (firms) is small but the number of observations per
entity is large. Their approach can be outlined as follows: Implement the estimation method (OLS, IV,
ITT) on each firm separately and obtain 11 firm-specific estimates. Note that we cannot do this for the
control firm since there is no within-firm variation for the right hand side for the control firms.
Therefore the results from this procedure are essentially based on before-after comparisons for the
treatment firms.

The procedure requires that the coefficient estimates from each entity are asymptotically independent
and Gaussian (but with potentially different variances). In our case this would be justified by an
asymptotics in T argument (recall we have about 120 observations per plant). In particular, we can be
agnostic about the exact structure of correlations between observations within a firm as long as the
parameter estimators satisfy a central limit theorem. Subject to this requirement, the extent of
correlation across observations within an entity is unrestricted. In addition, different correlation
structures across firms are permissible since the procedure allows for different variances for each firm
level parameter. This “asymptotic heterogeneity” considerably relaxes the usual assumptions made in
standard panel data contexts (such as those underlying the cluster covariance matrices in our main
tables). Finally, IM show that the limiting standard Gaussian distribution assumption (for each firm)
can be relaxed to accommodate heterogeneous scale mixtures of standard normal distributions as well.

The procedure comprises estimating separate coefficients for each of the 11 treatment firms (recall that
we cannot use control firms for this procedure since they have no within firm variation in the
instrument) and treating the resulting estimates as a set of independent random variables drawn from
normal distributions.

We next summarize the ideas underlying the permutation based tests.We first describe the permutation
test for the ITT parameter. We base the test on the Wei-Lachin (1994) statistic as described in Greevy
et al (2004). The reason for using this statistic (as opposed to more commonly used ones) is that the
permutation test for the IV parameter is a generalization of this procedure and so it natural to consider
this procedure in the first step. Consider the vector of outcomes for plant i (we examine each
outcome separately). Define the binary random assignment variable for firm i . Define the random
variable

This variable takes on the values 0, 1 and -1. It is equal to zero if plant is a control or plant is a
treatment plant of any of the outcome variables for either plant is missing. It is equal to +1 if plant is
a treatment plant, plant is a control and the outcome for i is larger than the outcome for j. It is equal
to -1 if plant is a treatment plant, plant is a control and the outcome for i is smaller than the
outcome for j. The Wei-Lachin statistic can be written as

Under the null hypothesis of no treatment effect, the treatment outcomes should not be systematically
larger than the control outcomes (note that this means that we do not use pre-treatment observations
for all plants). Specifically, under the null hypothesis and conditional upon the order statistics, each
possible candidate value of T has an an equal probability of occurring. We use this insight to construct
a critical value for the test. Consider one of the combinations of the firm treatment assignment
variable Z. For each such permutation, compute T. Form the empirical distribution of T by

32
considering all possible permutations and record the appropriate quantile for the distribution of T thus
generated (in the one-sided alternative case this would be the 1−α quantile). Finally, reject the null
hypothesis of no treatment effect if the original statistic T exceeds this quantile. Greevy et al (2004),
using previous results (Rosenbaum (2002) and Imbens and Rosenbaum (2005)) show that this test has
exact size α for any sample size n. Therefore, the conclusions of this test do not rely upon any
asymptotic theory. Instead, the results lean heavily on the idea of exchangeability – the property that
changing the ordering of a sequence of random variables does not affect their joint distribution. For
our application, this notion seems reasonable. Note that exchangeability is weaker than the i.i.d.
assumption so for instance outcomes across firms can even be correlated (as long as they are equi-
correlated).

Consider next the randomization inference based test for the IV case. We first consider the cross-
section. Define the counterfactual model for outcomes and let denote potential
treatment status when treatment assignment is . Define observed treatment status as
. In our case, the treatment status is the fraction of the 38 practies that the
firm has implemented. The maintained assumption is that the potential outcomes are independent of
the instrument Z or equivalently is independent of Z and the error term has mean 0. We
observe a random sample on and wish to test the null hypothesis against the
two-sided alternative. Note that under the null hypothesis, is independent of
Z and we use this fact to construct a test along the lines of the previous test. Consider the analogue of
the first equation

Where we have replaced the response by the response subtracted by . Note that is
consistently estimable under the null, so without loss of generality we can treat it as known. This point
is discussed in greater detail below. The key to this rewriting is to notice that under the null
hypothesis, is independent of Z so that we can use the same argument as before to derive the finite
sample distribution of T .For our data, we modify this approach to allow for a panel and covariates
(time and plant dummies). This parallels the proposal in Andrews and Marmer (2008) and we can
define

and we form the statistic as

Where

For each candidate value of , we form and carry out the permutation test (as described in the
ITT case above and noting that we do not use pre-treatment outcomes). We collect the set of values for
which we could not reject the null hypothesis (against the two-sided alternative at α=.05) to construct
an exact confidence set for . As Imbens and Rosenbaum (2005) point out, there is no reason for a
confidence set constructed in this manner to be a single interval. However, in our estimation, the
confidence sets were intervals.

33
Table A1: The textile management practices adoption rates
Area Specific practice Pre-intervention level Post-intervention change
Treatment Control Treatment Control
Preventive maintenance is carried out for the machines 0.429 0.667 0.286 0
Preventive maintenance is carried out per manufacturer's recommendations 0.071 0 0.071 0.167
The shop floor is marked clearly for where each machine should be 0.071 0.333 0.214 0.167
The shop floor is clear of waste and obstacles 0 0.167 0.214 0.167
Machine downtime is recorded 0.571 0.667 0.357 0
Machine downtime reasons are monitored daily 0.429 0.167 0.5 0.5
Factory Machine downtime analyzed at least fortnightly & action plans implemented to try to reduce this 0 0.167 0.714 0
Operations Daily meetings take place that discuss efficiency with the production team 0 0.167 0.786 0.5
Written procedures for warping, drawing, weaving & beam gaiting are displayed 0.071 0.167 0.5 0
Visual aids display daily efficiency loomwise and weaverwise 0.214 0.167 0.643 0.167
These visual aids are updated on a daily basis 0.143 0 0.643 0.167
Spares stored in a systematic basis (labeling and demarked locations) 0.143 0 0.143 0.167
Spares purchases and consumption are recorded and monitored 0.571 0667 0.071 0.167
Scientific methods are used to define inventory norms for spares 0 0 0.071 0
Quality defects are recorded 0.929 1 0.071 0
Quality defects are recorded defect wise 0.286 0.167 0.643 0.833
Quality defects are monitored on a daily basis 0.286 0.167 0.714 0.333
Quality There is an analysis and action plan based on defects data 0 0 0.714 0.167
Control There is a fabric gradation system 0.571 0.667 0.357 0
The gradation system is well defined 0.500 0.5 0.429 0
Daily meetings take place that discuss defects and gradation 0.071 0.167 0.786 0.167
Standard operating procedures are displayed for quality supervisors & checkers 0 0 0.714 0
Yarn transactions (receipt, issues, returns) are recorded daily 0.929 1 0.071 0
The closing stock is monitored at least weekly 0.214 0.167 0.571 0.5
Inventory Scientific methods are used to define inventory norms for yarn 0 0 0.083 0
Control There is a process for monitoring the aging of yarn stock 0.231 0 0.538 0
There is a system for using and disposing of old stock 0 0 0.615 0.6
There is location wise entry maintained for yarn storage 0.357 0 0.357 0
Loom Advance loom planning is undertaken 0.429 0.833 0.214 0
Planning There is a regular meeting between sales and operational management 0.429 0.500 0.143 0
There is a reward system for non-managerial staff based on performance 0.571 0.667 0.071 0
There is a reward system for managerial staff based on performance 0.214 0.167 0.286 0
Human
There is a reward system for non-managerial staff based on attendance 0.214 0.333 0.357 0
Resources
Top performers among factory staff are publicly identified each month 0.071 0 0.357 0
Roles & responsibilities are displayed for managers and supervisors 0 0 0.643 0
Customers are segmented for order prioritization 0 0 0 0.167
Sales and
Orderwise production planning is undertaken 0.692 1 0.231 0
Orders
Historical efficiency data is analyzed for business decisions regarding designs 0 0 0.071 0
All Average of all practices 0.256 0.288 0.378 0.120
p-value for the difference between the average of all practices 0.510 0.000
Notes: Reports the 38 individual management practices measured before, during and after the management intervention. The columns Pre Intervention level of Adoption report the pre-
intervention share of plants adopting this practice for the 14 treatment and 6 control plants. The columns Post Intervention increase in Adoption report the changes in adoption rates
between the pre-intervention period and 4 months after the end of the diagnostic phase (so right after the end of the implementation phase for the treatment plants) for the treatment and
control plants. The p-value for the difference between the average of all practices reports the significance of the difference in the average level of adoption and the increase in adoption
between the treatment and control groups.

34
 Table A2: The decentralization survey:
For all questions except D7 any score can be given, but the scoring guide is only provided for scores of 1, 3 and 5.
Question D1: “What authority does the plant manager(or other managers) have to hire a WEAVER (e.g. a worker supplied by a contractor)?”
Score 1 Score 3 Score 5
Scoring grid: No authority – even for replacement hires Requires sign-off from the Director based on the business case. Complete authority – it is my decision
Typically agreed (i.e. about 80% or 90% of the time). entirely

Question D2: “What authority does the plant manager(or other managers) have to hire a junior Manager (e.g. somebody hired by the firm)?”
Score 1 Score 3 Score 5
Scoring grid: No authority – even for replacement hires Requires sign-off from the Director based on the business case. Complete authority – it is my decision
Typically agreed (i.e. about 80% or 90% of the time). entirely

Question D3: “What authority does the plant manager (or other managers) have to purchase spare parts?”?
Probe until you can accurately score the question. Also take an average score for sales and marketing if they are taken at different levels.
Score 1 Score 3 Score 5
Scoring grid: No authority Requires sign-off from the Director based on the business case. Complete authority – it is my decision
Typically agreed (i.e. about 80% or 90% of the time). entirely

Question D4: “What authority does the plant manager (or other managers) have to plan maintenance schedules?”
Score 1 Score 3 Score 5
Scoring grid: No authority Requires sign-off from the Director based on the business case. Complete authority – it is my decision
Typically agreed (i.e. about 80% or 90% of the time). entirely

Question D5: “What authority does the plant manager (or other managers) have to award small (<10% of salary) bonuses to workers?”
Score 1 Score 3 Score 5
Scoring grid: No authority Requires sign-off from the Director based on the business case. Complete authority – it is my decision
Typically agreed (i.e. about 80% or 90% of the time). entirely

Question D6: “What is the largest expenditure (in rupees) a plant manager (or other managers) could typically make without your signature?”

Question D7: “What is the extent of follow-up required to be done by the directors?”
Score 1 Score 3 Score 5
Scoring grid: Directors are the primary point of contact for Frequent follow ups on about half of the decisions made by Minimal follow-ups on decisions taken
information exchange between managers managers between managers. Only dispute
resolution.
Question D8: “How many days a week did the director spend away from the factory last month?”

35
Table A3: The computerization survey:
Question C1: “Does the plant have an Electronic resource planning system?”
Question C2: “How many computers does the plant have?”
Question C3: “How many of these computers are less than 2 years old”
Question C4: “How many people in the factory typically use computers for at least 10 minutes day?”
Question C5: “How many cumulative hours per week are computers used in the plant”?
Question C6: “Does the plant have an internet connection”
Question C7: “Does the plant manager use e-mail (for work purposes)?”
Question C8: “Does the plant manager use e-mail (for work purposes)?”
Question C9: “What is the extent of computer use in operational performance management?” (and score from 1 to 5 is possible, but scores given for 1,3, and 5)
Score 1 Score 3 Score 5
Scoring grid: Computers not used in Around 50% of operational performance metrics All main operational performance metrics (efficiency,
operational performance (efficiency, inventory, quality and output) are tracked inventory, quality and output) are tracked & analyzed
management & analyzed through computer/ERP generated reports. through computer/ERP generated reports.

Table A4: Descriptive statistics for the Decentralization and Computerization survey
Mean Min pre- Max SD Mean Correlation of change
pre-level level pre-level pre-level change with treatment status
Decentralization questions
D1 (weaver hiring) 4.68 3 5 0.72 0 n/a
D2 (manager hiring) 1.93 1 4 1.05 0.36 0.198
D3 (spares purchases) 2.61 1 4 0.79 0.18 0.121
D4 (maintenance planning) 4.50 1 5 1 0.04 0.133
D5 (worker bonus pay) 2.25 1 4 1.14 0.29 0.375
D6 (investment limit, rupees) 10357 1000 35000 10434 714 0.169
D7 (director coordination) 2.78 2 4 0.69 0.36 0.358
D8 (days director not at the factory per week) 2.69 0 4.75 1.30 0.39 0.282
Decentralization index 0 -1.33 1.52 1 0.44 0.355
Computerization questions
C1 (ERP) 0.74 0 1 0.44 0 n/a
C2 (number computers) 2.68 0 8 1.98 0.36 0.377
C3 (number new computers) 0.43 0 8 1.55 0.29 0.189
C4 (computer users) 3 0 10 2.21 0.11 0.308
C5 (computer hours) 10 0 48 12.20 5.34 0.439
C6 (internet connection) 0.64 0 1 0.49 0.036 0.133
C7 (plant manager e-mail) 0.29 0 1 0.46 0.04 -0.280
C8 (directors e-mail) 0.82 0 1 0.39 0 n/a
C9 (production computerization) 2.71 1 5 0.98 0.89 0.367
Computerization index 0 -1.58 3.15 1 0.458 0.440
Notes: There are about 50 rupees to the dollar. The mean change measures the different between pre the experiment and August 2010. The decentralization index and the
computerization index are normalized to have a mean of zero and standard deviation of unity on the pre-experiment sample.

36
Table A5: Estimated impact of improved quality, inventory and efficiency

Change Impact Estimation approach Estimated

impact
Profits (annual in $)
Improvement in Reduction in Reduction in defects (32%) times median $13,000
quality repair mending manpower wage bill ($41,000).
manpower
Reduction in Reduction in defects (32%) times the $96,000
waste fabric average yearly waste fabric (5%) times
median average sales ($6m).

Reduction in Reduction in Reduction in inventory (16.4%) times $8,000

inventory inventory carrying cost of inventory (22%) times
carrying costs median inventory ($230,000)

Increased Increased sales Increase in output (5.4%) times margin $121,000

efficiency on sales (37% ) times median sales ($6m)
Total $238,000
Productivity (%)
Improvement in Reduction in Reduction in defects (32%) times share 3.5%
quality repair of repair manpower in total manpower
manpower (18.7%) times labor share (0.58) in
output in textiles (from the 2003-04
Indian Annual Survey of Industries.)

Reduction in Reduction in defects (31.9%) times the 1.6%

waste fabric average yearly waste fabric (5%)

Reduction in Reduction in Reduction in inventory (16.4%) times 0.6%

inventory capital stock inventory share in capital (8%) times
capital factor share in output in textiles
(0.42)

Increased Increased Increase in output (5.4%) without any 5.4%

efficiency output change in labor or capital
Total 11.1%
Notes: Estimated impact of the improvements in the management intervention on firms profitability and
productivity through quality, inventory and efficiency using the estimates in Table 3. Figure calculated for the
median firm. See Appendix A for details of calculations for inventory carrying costs, fabric waste, repair
manpower and factor shares.

37
 Table 1: The field experiment sample

All Treatment Control Diff

Mean Median Min Max Mean Mean p-value
Sample sizes:
Number of plants 28 n/a n/a n/a 19 9 n/a
Number of experimental plants 20 n/a n/a n/a 14 6 n/a
Number of firms 17 n/a n/a n/a 11 6 n/a
Plants per firm 1.65 2 1 4 1.73 1.5 0.393
Firm/plant sizes:
Employees per firm 273 250 70 500 291 236 0.454
Employees, experimental plants 134 132 60 250 144 114 0.161
Hierarchical levels 4.4 4 3 7 4.4 4.4 0.935
Annual sales $m per firm 7.45 6 1.4 15.6 7.06 8.37 0.598
Current assets $m per firm 12.8 7.9 2.85 44.2 13.3 12.0 0.837
Daily mtrs, experimental plants 5560 5130 2260 13000 5,757 5,091 0.602
Management and plant ages:
BVR Management score 2.60 2.61 1.89 3.28 2.50 2.75 0.203
Management adoption rates 0.262 0.257 0.079 0.553 0.255 0.288 0.575
Age, experimental plant (years) 19.4 16.5 2 46 20.5 16.8 0.662
Performance measures
Operating efficiency (%) 70.77 72.8 26.2 90.4 70.2 71.99 0.758
Raw materials inventory (kg) 59,497 61,198 6,721 149,513 59,222 60,002 0.957
Quality (% A-grade fabric) 40.12 34.03 9.88 87.11 39.04 41.76 0.629

Notes: Data provided at the plant and/or firm level depending on availability. Number of plants is the total
number of textile plants per firm including the non-experimental plants. Number of experimental plants is the
total number of treatment and control plants. Number of firms is the number of treatment and control firms.
Plants per firm reports the total number of other textiles plants per firm. Several of these firms have other
businesses – for example retail units and real-estate arms – which are not included in any of the figures here.
Employees per firm reports the number of employees across all the textile production plants, the corporate
headquarters and sales office. Employees per experiment plant reports the number of employees in the
experiment plants. Hierarchical levels displays the number of reporting levels in the experimental plants – for
example a firm with workers reporting to foreman, foreman to operations manager, operations manager to the
general manager and general manager to the managing director would have 4 hierarchical levels. BVR
Management score is the Bloom and Van Reenen (2007) management score for the experiment plants.
Management adoption rates are the adoption rates of the management practices listed in Table 2 in the
experimental plants. Annual sales ($m) and Current assets ($m) are both in 2009 US $million values,
exchanged at 50 rupees = 1 US Dollar. Daily mtrs, experimental plants reports the daily meters of fabric
woven in the experiment plants. Note that about 3.5 meters is required for a full suit with jacket and trousers, so
the mean plant produces enough for about 1600 suits daily. Age of experimental plant (years) reports the age
of the plant for the experimental plants. Note that none of the differences between the means of the treatment and
control plants are significant. Raw materials inventory is the stock of yarn per intervention. Operating
efficiency is the percentage of the time the machines are producing fabric per intervention. Quality (% A-grade
fabric) is the percentage of fabric each plant defines as A-grade, which is the top quality grade.

38
Table 2: The impact of the treatment on management practices within and across plants
Dependent Variable Overall Overall Overall Overall Overall
Management Management Management Management Management
(1) (2) (3) (4) (5)
Own plant treatmenti.t 0.109*** 0.108*** 0.108*** 0.105*** 0.105***
Months consulting in own plant (0.014) (0.016) (0.016) (0.018) (0.018)
Spillover treatmenti,t 0.041** -0.007 0.030
Months consulting in other plants within the same firm (0.016) (0.023) (0.024)

3 month lagged spillover treatmenti,t-3 0.054** 0.047***

Months consulting in other plants within the same firm (0.023) (0.016)
6 month lagged spillover treatmenti,t-6 0.027 0.050***
Months consulting in other plants within the same firm (0.021) (0.017)

Time FEs 12 11 11 10 10
Plant FEs 28 28 28 28 28
Observations 336 308 308 280 280
R-squared 0.904 0.909 0.909 0.889 0.889
Notes: The dependent variable is the share of the 38 management practices adopted in each plant. This is regressed against the cumulative weeks of intervention
in the own plant (“Own plant treatment”), the cumulative weeks of treatment in other plants within the same firm (“Spillover treatment”), and this variable
lagged three and six months (“3 month Lag spillover treatment” and “6 month Lag spillover treatment”). The data is quarterly until April 2009 and bi-monthly
thereafter, reflecting the frequency of measurement of management practices. A full set of time-dummies and plant dummies is included. Standard errors are
bootstrap clustered at the firm level.

39
Table 3: The impact of modern management practices on plant performance
Dependent Variable Quality (log QDI) Inventory (log tons) Output (log picks)
Specification OLS IV ITT OLS IV ITT OLS IV ITT
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Managementi,t -0.561 -2.028*** -0.639*** -0.929** 0.127 0.346**
Adoption of management practices (0.440) (0.685) (0.242) (0.389) (0.099) (0.171)
Interventioni,t -0.386** -0.179** 0.056
Intervention stage initiated (0.162) (0.089) (0.034)

Log (1+ Log (1+ Log (1+

Instrument weeks weeks weeks
treatment) treatment) treatment)
Small sample robustness
Ibragimov-Mueller (95% CI) (-4.46,-0.53) (-5.03,-0.98) (-0.69,-0.38) (-0.81,-0.09) (-0.84,-0.02) (-0.17,0.02) (0.22,0.86) (-.08,2.25) (0.05,0.26)
(90%CI) (-4.09,-0.90) (-4.65,-1.36) (-0.66,-0.41) (-0.75,-0.16) (-0.77,-0.10) (-0.15,0.00) (0.28,0.80) (0.13,2.03) (0.07,0.24)
Permutation Test I (p-value) .02 .14 .10
IV Permutation Tests (95% CI) (-6.05,-.06) (-2.64,0.66) (0.07,0.62)
(90% CI) (-6.00,-0.21) (-2.04,0.52) (0.12,0.54)
Time FEs 113 113 113 113 113 113 114 114 114
Plant FEs 20 20 20 18 18 18 20 20 20
Observations 1732 1732 1732 1977 1977 1977 2312 2312 2312
Notes: All regressions use a full set of plant and calendar week dummies. Standard errors bootstrap clustered at the firm level. Quality (log QDI) is a log of the
quality defects index (QDI), which is a weighted average score of quality defects, so higher numbers imply worse quality products (more quality defects).
Inventory (log tons) is the log of the tons of yarn inventory in the plant. Output (log picks) is the log of the weaving production picks. Management is the
adoption of the 38 management practices listed in table 2. Intervention (implementation) is a plant level indicator taking a value of 1 after the implementation
phase has started at a treatment plant. Log(1+weeks of treatment) is the log of one plus the cumulative count of the weeks since the start of the implementation
in each plant (treatment plants only), and value zero before. OLS reports results with plant estimations. IV reports the results where the management variable has
been instrumented with log(1+ cumulative intervention weeks). ITT reports the intention to treat results from regressing the dependent variable directly on the
1/0 intervention indicator. Time FEs report the number of calendar week time fixed effects. Plant FEs reports the number of plant-level fixed effects. Two
plants do not have any inventory on site, so no inventory data is available. The Small sample robustness implements three different procedures (described in
greater detail in Appendix B) to address issues of plant heterogeneity, within plant (and firm) correlation, and small sample concerns, where 95% CI and 90%
CI report 95% and 90% confidence intervals. Ibragimov-Mueller estimates parameters firm-by-firm and then treats the estimates as a draw from independent
(but not identically distributed) normal distributions. Permutation Test I reports the p-values for testing the null hypothesis that the treatment has no effect for
the ITT parameter by constructing a permutation distribution of the ITT estimate using 1000 possible permutations (out of 12376) of treatment assignment. IV-
Permutation tests implements a permutation test for the IV parameter using 1000 possible permutations (out of 12376) of treatment assignment. These tests have
exact finite sample size.

40
Table 4: The impact of modern management practices on organization and computerization
Dep. variable: Decentralization Index Hours of computer use Computerization index
Specification OLS IV ITT OLS IV ITT OLS IV ITT
(1) (2) (7) (4) (5) (6) (7) (8) (9)
Managementi,t 1.695*** 1.837*** 16.761*** 23.272** 1.154*** 1.497**
Adoption of management practices (0.420) (0.535) (3.457) (6.708) (0.338) (0.555)
Interventioni,t 0.360** 6.168** 0.403**
Intervention stage initiated (0.164) (2.163) (0.148)
Log Log Log
Instrument (1+weeks (1+weeks (1+weeks
treatment) treatment) treatment)
Small sample robustness
Ibragimov-Mueller (95% CI)
(90%CI)
Permutation Test I (p-value)
IV Permutation Tests (95% CI)
(90% CI)
Time FEs 3 3 3 3 3 3 3 3 3
Plant FEs 28 28 28 28 28 28 28 28 28
Observations 84 84 84 84 84 84 84 84 84
Notes: All regressions use three observations per firm (pre intervention, March 2010 and August 2010), and a full set of plant dummies and time dummies.
Standard errors bootstrap clustered at the firm level. Management is the adoption of the 38 management practices listed in table A1. Decentralization index is
the principal component factor of 7 measures of decentralization around weaver hiring, manager hiring, spares purchases, maintenance planning, weaver
bonuses, investment, and departmental co-ordination. This has a standard deviation of 1 and a mean of 0. Hours of computer use is the hours of computer use.
This has a (pre-intervention) mean and standard deviation of 13.66 and 12.20. Computerization index is the principal component factor of 10 measures around
computerization, which are the use of an ERP system, the number of computers in the plant, the number of computers less than 2 years old, the number of
employees using computers for at least 10 minutes per day, and the cumulative number of hours of computer use per week, an internet connection at the plant, if
the plant-manager uses e-mail, if the directors use of e-mail, and the intensity of computerization in production. The other computerization columns show the
results for the individual components of this index that changed over time (the omitted components did not change). This has a standard deviation of 1 and a
mean of 0. Log(1+weeks of treatment) is the log of one plus the cumulative count of the weeks since the start of the implementation in each plant (treatment
plants only), and value zero before. OLS reports results with plant estimations. IV reports the results where the management variable has been instrumented with
log(1+ cumulative intervention weeks). ITT reports the intention to treat results from regressing the dependent variable directly on the 1/0 intervention indicator.
Time FEs reports the number of time fixed effects. Plant FEs reports the number of plant-level fixed effects. SD of dep. var. reports the standard deviation of
the dependent variable. The Small sample robustness implements three different procedures (described in greater detail in Appendix B) to address issues of
plant heterogeneity, within plant (and firm) correlation, and small sample concerns.

41
 Table 5: Reasons for bad management, as a percentage (%) of all practices, before and after treatment
Non-adoption reason Firm group 1 1 3 5 7 9
month month months months months months
before after after after after after
Treatment 18.5 13.5 2.0 0.6 0 0
Lack of information
Control 12.9 9.6 8.0 8.0 8.0 8.0
(plants not aware of the practice)
Non-experimental 9.3 6.8 3.8 3.8 3.8 3.8
Treatment 44.4 36.6 33.6 31.3 31.1 30.2
Incorrect information
Control 46.7 45.3 44.2 43.1 42.2 42.2
(plants incorrect on cost-benefit calculation)
Non-experimental 41.2 42.0 38.6 35.6 34.6 33.6
Treatment 10.3 7.5 7.2 7.5 7.7 6.8
Owner lack of time, low ability or procrastination
Control 11.6 10.2 9.3 9.8 8.4 8.4
(the owner is the reason for non adoption)
Non-experimental 23.5 22.0 27.0 31.5 26.3 26.6
Treatment 0.5 0.5 0.5 0.5 0.5 0.5
Not profitable
Control 0 0 0 0 0 0
(the consultants agree non-adoption is correct)
Non-experimental 0 0 0 0 0 0
Treatment 0.2 0.2 0.2 0.2 0.2 0.2
Other
Control 0 0 0 0 0.9 0.9
(variety of other reasons for non-adoption)
Non-experimental 0.3 0.3 0.3 0.3 0.3 0.3
Total Treatment 74.4 58.2 45.5 40.1 39.9 38.1
(sum of all individual reasons) Control 71.2 65.1 61.6 60.9 60.6 60.6
Non-experimental 73.4 71.0 70.7 69.8 65.4 64.7

Notes: Show the percentages (%) of practices not adopted by reason for non-adoption, in the treatment plants, control plants and non-experimental plants.
Timing is relative to the start of the treatment phase (the end of the diagnostic phase for the control group and the start of the treatment phase for the other plant
in their firm for the non-experimental plants). Covers 532 practices in treatment plants (38 practices in 14 plants), 228 practices in the control plants (38 practices
in 6 plants) and 30 practices in the non-experimental plants (38 practices in 8 plants). Non adoption was monitored every other month using the tool shown in
Figure 4, based on discussions with the firms’ directors, managers, workers, plus regular consulting work in the factories. Note that data is only currently
available up to 7 months after the end of diagnostic phase in the control firms.

42
Exhibit 1: Plants are large compounds, often containing several buildings.

Plant entrance with gates and a guard post Plant surrounded by grounds

Front entrance to the main building Plant buildings with gates and guard post

Exhibit 2: These factories operate 24 hours a day for 7 days a week

producing fabric from yarn, with 4 main stages of production

(1) Winding the yarn thread onto the warp beam (2) Drawing the warp beam ready for weaving

(3) Weaving the fabric on the weaving loom (4) Quality checking and repair
Exhibit 3: Many parts of these factories were dirty and unsafe

Garbage outside the factory Garbage inside a factory

Flammable garbage in a factory Chemicals without any covering

Exhibit 4: The factory floors were frequently disorganized

Instrument
not Old warp
removed beam, chairs
after use, and a desk
blocking obstructing the
hallway. factory floor

Dirty and
poorly Tools left on
maintained the floor
machines after use
Exhibit 5: Most plants had months of excess yarn, usually spread across
multiple locations, often without any rigorous storage system

Yarn without Yarn piled up so high and

labeling, order or deep that access to back
damp protection sacks is almost impossible

Different types
and colors of Crushed yarn cones (which
yarn lying mixed need to be rewound on a
new cone) from poor storage

Exhibit 6: The parts stores were often disorganized and dirty

Spares without any labeling or No protection to prevent damage and rust

order

Spares without any labeling or order Shelves overfilled and disorganized

Exhibit 7: Non adoption flow chart used by consultants to collect data
Legend Is the reason for the non adoption  No
of the practice internal to the firm? External factors (legal, climate etc)
Hypothesis
Yes
Conclusion
Was the firm previously aware
Yes Lack of information
that the practice existed?

No

Could the firm hire new 
Can the firm adopt the practice with  employees or consultants  Lack of local skills
existing staff & equipment? to adopt the practice?

Did the owner believe introducing  Would this adoption be 
profitable Not profit maximizing
the practice would be profitable?

Owner lack of time, low Could the CEO get his employees to  Do you think the CEO was correct 

ability or procrastination introduce the practice? about the cost‐benefit tradeoff?

Did the firm 
Does the firm have enough internal  realize this 
Incorrect information
financing or access to credit?   would be 
profitable? 

Other reasons Credit constraints

Notes: The consultants used the flow chart to evaluate why each particular practice from the list of 38 in Table 2 had not been
adopted in each firm, on a bi-monthly basis. Non adoption was monitored every other month based on discussions7with the firms’
directors, managers, workers, plus regular consulting work in the factories.
 Figure 1: Management practice scores across countries
8
.

US Manufacturing, mean=3.33
0
06
.
4
.
2
.

1
1 2 3
3 4 5
5
8
.
6
.

Indian Manufacturing, mean=2.69

s
ny

4
D
et
i

.
2
.
0 0

1 1 2 33 4 5
5
Density of Firms
8
.
6

Brazil and China Manufacturing, mean=2.67

.
s
n
D
ey
t
i

0
0 4
.
2
.

1
1 2 3
3 4 5
5
1
8
.
6

Indian Textiles, mean=2.60

y

.
s
n
D
et
i

4
.
2
.
0

1 2 3 4 5
5
1
.

Experimental Firms, mean=2.60

1
s
n
D
ey
t
i

0
0 5
.

1
1 3 5
5

Management score
Notes: Management practice histograms using the Bloom and Van Reenen (2010) data. Double-blind surveys used to evaluate
firms monitoring, targets and operations. Scores range from 1 (worst practice) to 5 (best practice). Samples are 17 experimental
plants, 232 Indian textile firms, 620 Indian manufacturing firms, and 1083 Brazilian and Chinese firms.
 Figure 2: The adoption of key textile management practices over time
Share of key textile management practices adopted
.6

Treatment plants
.5

Control plants
.4
.3

Non-experimental plants
.2

-10 -8 -6 -4 -2 0 2 4 6 8 10 12
Months after the diagnostic phase
Notes: Average adoption rates of the 38 key textile manufacturing management practices listed in Table 2. Shown separately for
the 14 treatment plants (round symbol), 6 control plants (diamond symbol) and the 8 non-experimental plants which the
consultants did not provide any direct consulting assistance to (+ symbol). Scores range from 0 (if none of the group of plants
have adopted any of the 38 management practices) to 1 (if all of the group of plants have adopted all of the 38 management
practices). Initial differences across all the groups are not statistically significant.
 Figure 3: Quality defects index for the treatment and control plants
Start of Start of End of
Diagnostic Implementation Implementation
Quality defects index (higher score=lower quality)

140

97.5th percentile
120
100

Average (♦ symbol)
Control plants
80

2.5th percentile
60

97.5th percentile

Average (+ symbol)
40

Treatment plants

2.5th percentile
20
0

-20 -10 0 10 20 30 40 50
Weeks after the start of the diagnostic
Notes: Displays the average weekly quality defects index, which is a weighted index of quality defects, so a higher score means
lower quality. This is plotted for the 14 treatment plants (+ symbols) and the 6 control plants (♦ symbols). Values normalized so
both series have an average of 100 prior to the start of the intervention. To obtain confidence intervals we bootstrapped the firms
with replacement 250 times.
 Figure 4: Yarn inventory for the treatment and control plants
Start of Start of End of
Diagnostic Implementation Implementation
97.5th percentile
120
Yarn inventory (normalized to 100 prior to diagnostic)

Average (♦ symbol)

Control plants
100

97.5th percentile

2.5th percentile

Average (+ symbol)

Treatment plants
80

2.5th percentile
60

-20 -10 0 10 20 30 40 50
Weeks after the start of the intervention
Notes: Displays the weekly average yarn inventory plotted for 12 treatment plants (+ symbols) and the 6 control plants (♦
symbols). Values normalized so both series have an average of 100 prior to the start of the intervention. To obtain confidence
intervals we bootstrapped the firms with replacement 250 times. Slow moving fluctuations due to seasonality. 2 treatment plants
maintain no on-site yarn inventory so are not included in the figures.
 Figure 5: Output for the treatment and control plants
Start of Start of End of
Diagnostic Implementation Implementation
130

97.5th percentile
120
Output (normalized to 100 prior to diagnostic)

Treatment plants
Average (+ symbol)
110

2.5th percentile
100

97.5th percentile

Average (♦ symbol)
90

Control plants

2.5th percentile
80
70

-20 -10 0 10 20 30 40 50
Weeks after the start of the intervention
Notes: Displays the weekly average output for the 14 treatment plants (+ symbols) and the 6 control plants (♦ symbols). Values
normalized so both series have an average of 100 prior to the start of the intervention. To obtain confidence intervals we
bootstrapped the firms with replacement 250 times.
Figure 6: Plant level changes in performance
Control Treatment Control Treatment

20
8

15
6
Density

Density
10
4

5
2

0
0

-1 -.5 0 .5 1 -1 -.5 0 .5 1 -.4 -.2 0 .2 .4 -.4 -.2 0 .2 .4

Before/after difference in log (Inventory)
Before/after difference in log (Quality Defects Index)

Notes: Displays the histogram of plant by plant changes

Control Treatment
in log (Quality Defects Index), log (Inventory) and log
(Output) between the post and pre treatment periods.
15
10
Density
5
0

0 .2 .4 .6 0 .2 .4 .6
Before/after difference in log (Output)