# Exporting and Firm Performance: Evidence from a Randomized Experiment

**David Atkin, Amit K. Khandelwal, and Adam Osman**

*Quarterly Journal of Economics, 2017, 132(2): 551-615*

---


}


## Abstract

We conduct a randomized experiment that generates exogenous variation in
the access to foreign markets for rug producers in Egypt.
Combined with detailed survey data, we causally identify the impact of
exporting on firm performance.
Treatment firms report 16-26 percent higher profits and exhibit large improveme
nts in quality alongside reductions in output per hour relative to control
firms.
These findings do not simply reflect firms being offered higher margins
to manufacture high-quality products that take longer to produce.
Instead, we find evidence of learning-by-exporting whereby exporting improves
technical efficiency.
First, treatment firms have higher productivity and quality after controlling
for rug specifications.
Second, when asked to produce an
*identical*
domestic rug using the same inputs and same capital equipment, treatment
firms produce higher quality rugs despite no difference in production time.
Third, treatment firms exhibit learning curves over time.
Finally, we document knowledge transfers with quality increasing most along
the specific dimensions that the knowledge pertained to.


## Abstract

**Keywords**
: Exports, Quality, Learning-by-Exporting, Productivity, Market Access


## Abstract

**JEL Codes**
: F10, F14, D24


## Introduction


There are large differences in productivity across countries (
Hall and Jones (1999)

,
Bloom and Van Reenen (2007)

).
The belief that access to high-income markets can help firms in developing
countries close this gap is one motivation behind the large resources now
flowing to market access initiatives.
For example, the WTO's Aid-for-Trade Initiative secured $48 billion in
annual commitments to help developing countries overcome “trade-related
constraints”, and the last two decades have seen a tripling in the


number of national export-promotion agencies that help domestic firms match
with foreign buyers (

).


Central to these programs achieving this goal is the belief that exporting
improves the productivity of firms, a mechanism called learning-by-exporting
(
Clerides et al. (1998)

,
Loecker (2007)

,
Harrison and Rodriguez-Clare (2010)

).
In the presence of learning-by-exporting, trade generates efficiency gains
which narrow this productivity gap and magnify the gains from trade relative
to models without learning such as those explored by
ACR2012

(e.g.
see
Alvarez et al. (2013)

).

Despite the pervasiveness of these initiatives, there is still an ongoing
debate as to whether exporting has a causal impact on measures of firm
performance.
Moreover, if performance does improve, it is unclear whether such improvements
occur through learning-by-exporting—outward shifts in the production possibilit
y frontier (PPF)—or simply through movements along the PPF.
There are two central challenges in answering these questions.
First, more productive firms select into exporting (for example, see
MelitzEMA2003

).
This selection has plagued empirical attempts to identify the causal impact
of exporting on firm performance because what appears to be higher productivity
among exporters may simply be self-selection.
The second difficulty is that researchers typically lack detailed information
required to isolate changes that occur within firms due to exporting.
The literature commonly uses revenue-based total factor productivity (TFPR)
measures which also reflect changes in markups, the product mix, and product
quality (
Loecker (2014)

).
This is problematic for identifying learning-by-exporting since the internation
al trade literature suggests all three are likely to change with exporting.
While quantity-based total factor productivity measures (TFPQ) solve problems
related to changing markups, standard datasets do not provide the level
of detail required to account for changes in product specifications or
quality.
[^1]

Hence, if trade causes firms to change along these dimensions, measured
improvements in quantity-based productivity measures may simply reflect
movements along the PPF, rather than outward shifts of the PPF.

This paper conducts a randomized control trial (RCT) on rug manufacturers
in Egypt to examine how exporting affects profits and productivity.
To our knowledge, this is the first attempt to generate exogenous firm-level
variation in exporting.
As explained in detail below, we achieved this through an intervention
that reduced matching frictions between foreign buyers and a random subset
of Egyptian firms.
Using this experimental variation, we uncover if and how an economic primitive-
–firm productivity—responds to exporting.

The random assignment into exporting directly addresses the first of the
two challenges detailed above: selection of firms into exporting.
We provided a subset of firms with the opportunity to export handmade carpets
to high-income markets.
To provide this opportunity, we partnered with a US-based non-governmental
organization (NGO) and an Egyptian intermediary to secure export orders
from foreign buyers through trade fairs and direct marketing channels.
With orders in hand, we surveyed a sample of several hundred small rug
manufacturers located in Fowa, Egypt.
A random subsample of these firms was provided with an initial opportunity
to fill the orders by producing 110m

of rugs (approximately 11 weeks of work).
As in a standard buyer-seller relationship, firms were offered subsequent
orders provided they were able to fulfill the initial orders to the satisfactio
n of the buyer and intermediary.
Prior to our study, only a limited number of firms had ever knowingly exported
their products.
Hence, we interpret our experimental design as providing non-exporting
firms with the opportunity to sell to high-income markets.

To address the second challenge in identifying the impact of exporting—measureme
nt—we tracked performance measures through periodic surveys of both treatment
firms (those who received the opportunity to export) and control firms
(those who received no such opportunity).
Our production-line level data allow us to record not just quantity data,
but also detailed specifications for the rugs being produced at the time
of each survey round.
These specifications include product categories within the flatweave-rug
segment and attributes, such as the thread count, which a buyer chooses
when the order is placed.
This level of detail allows us to control for changes in the product mix
due to exporting with much more accuracy than is possible in typical datasets
(e.g.
using HS-10 product codes in trade datasets).
To further guarantee we are not conflating changes in productivity with
changes in the product mix, at the end of the study all firms were paid
to make an identical domestic-market rug using the same inputs and equipment.
To analyze changes in product quality with exporting, we collect direct
measures of product quality along 11 dimensions from a skilled quality
assessor who visited each firm in each survey round.
These quality measures capture a combination of both specifications and
hard-to-codify attributes that depend on the technical skill of the firm,
such as how flat the rug lies on the floor or how sharp the corners are.
Finally, we collect data on information flows between buyers, the intermediary
and producers that include transcripts of buyer feedback and the content
of discussions between the intermediary and the producers.
Taken together, these data allow us to address directly the measurement
challenges noted above.

Thanks to the randomization procedure, the causal effects of exporting are
identified by comparing mean outcomes between treatment and control firms.
We find that the opportunity to export raises the overall performance of
firms as measured by profits—treatment firms report 16-26 percent higher
profits relative to control firms.
The substantial increase in profits is perhaps not surprising given that
firms were provided with a positive demand shock, but is interesting given
the more moderate profit impacts the literature has found when exploring
supply-side interventions such as credit access (

).

The primary focus of this paper is to understand the mechanisms driving
the profit increases.
Despite increases in output prices and labor hours, we observe a decline
in total output (m

of rugs produced) among treatment firms.
These findings suggest that buyers from high-income countries demand higher-qua
lity rugs that take longer to produce.
Indeed, our quality assessments show that the rugs produced by treatment
firms score significantly higher along virtually every quality dimension.
At the same time,
"

unadjusted
"

productivity measures—those that do not control for rug specifications
and quality (e.g.
output per hour)—fall by 24-28 percent among treatment firms.

A simple theoretical framework shows that the quality upgrading we find
is consistent with two distinct mechanisms that have not been disentangled
in the literature to date.
The first mechanism amounts to a movement along the PPF.
We posit that the output per unit input of a firm depends on both rug specifica
tions and an output efficiency parameter
$\chi_{a}$

; high-specification rugs take longer to weave and,
*ceteris paribus*
, firms with higher
$\chi_{a}$

produce more output per unit input.
Quality also depends on rug specifications and a quality efficiency parameter,

$\chi_{q}$

, and is increasing in both.
The export opportunity exposes firms to buyers in high-income markets,
and these buyers are willing to pay more for quality than domestic buyers.
As long as firms find it profitable to do so, they will raise specifications,
and hence improve quality.
Under this first mechanism, firms already know how to manufacture high-quality
rugs and the opportunity to export simply induces a movement
*along*
the PPF.
That is, there is no change in either efficiency parameter.

A second mechanism involves an increase in the efficiency parameters induced
by exporting: learning-by-exporting.
This learning can come about through transfers of knowledge from buyers
to producers, or from learning-by-doing if such learning would not have
happened without exporting (a distinction we return to).
Learning-by-exporting is an

*outward*

shift of the PPF which can occur either by raising
$\chi_{a}$

(producing more output per input conditional on specifications) or raising

$\chi_{q}$

(producing higher quality conditional on specifications).
When these increases in efficiency are biased towards the production of
high-quality rugs, both rug quality and profits will rise.
Of course, the two mechanisms are not mutually exclusive, but the presence
of learning-by-exporting is important because it implies larger gains from
trade.

We present five pieces of evidence to show that the improvements in performance
come, at least in part, through learning-by-exporting.
The first is that both quality and productivity
*rise*
after adjusting for product specifications (recall that
"

unadjusted
"

productivity falls).
If firms only moved along the PPF, specification-adjusted quality and productiv
ity would remain constant.
Second, to ensure specifications were fully controlled for, at the end
of our experiment we asked all firms to manufacture an

*identical-specification*

rug for the domestic market using identical inputs and a common loom in
a workshop that we leased (a
"

quality lab
"

).
The rugs that treatment firms produce received higher scores along every
quality metric and were more accurate in terms of the desired size and
weight; moreover, treatment firms do not take longer


to produce these rugs despite their higher levels of quality.
Third, we explore the evolution of quality and productivity over time.
Inconsistent with a movement along the PPF, where quality and productivity
should immediately jump and then stay fixed, we document learning curves
for both.
Fourth, we draw on correspondences between foreign buyers and the intermediary,
as well as a log book of discussions between the intermediary and producers,
to document that our results come, in part, from knowledge flows.
In particular, we show that treatment firms improve quality most along
the particular quality dimensions discussed during meetings between the
intermediary and the producer, and that the vast majority of these discussions
were about specific weaving techniques to improve quality.
This, coupled with more anecdotal reports of knowledge passed from foreign
buyers to the intermediary, suggests that the improvements in efficiency
occur partly through knowledge transfers from intermediaries and foreign
buyers.
Fifth, we rule out investment explanations by showing that treatment firms
make no monetary or time investments in upgrading; nor do they pay, even
implicitly, for the knowledge they receive from the intermediary.
Taken together, the evidence strongly supports the presence of learning-by-expo
rting.

As with any industry- or country-specific study, it is important to acknowledge
issues relating to external validity.
In terms of the context of our study, the firms in our sample are small—typical
ly having only one full-time employee—and production is not automated.
Hence, our study has little to say about learning-by-exporting for large
firms manufacturing complex products.
Of course, it is precisely their small size that allows us to assemble
a large sample necessary for inference; and the fact that they manufacture
products using the same technology allows us to design more specific surveys
and improves statistical power.
The firms also export via an intermediary, rather than directly, but indirect
exporting is common for the Egyptian rug industry and in other industries
and countries.

In terms of the experiment itself, our treatment induces exporting by reducing
matching frictions between firms and sophisticated foreign buyers.
Such frictions are of interest both for theoretical and policy reasons.

Allen2014

estimates that matching frictions explain half of the overall variation
in trade costs, while
Lederman et al. (2010)

note that reducing matching frictions for small- and medium-sized firms
(SMEs) is a key goal for export promotion agencies.
That said, our experiment does not reduce the trade frictions more typically
studied in the trade literature such as tariffs or transport costs.
While ultimately an empirical question requiring further research, we conjectur
e that exporting to high-income markets will generate similar learning for
developing-country SMEs however they are induced to export (whether by
reducing tariffs, trade costs or matching frictions).

Two more caveats are necessary.
Given the difficulties we document in generating orders and the implicit
labor costs of our time, it is unclear if it is efficient for our sample
firms to pay the fixed costs required to find sophisticated foreign buyers.
The goal of this paper is not to carry out a cost-benefit analysis of export
facilitation programs, or to isolate market failures preventing firms from
exporting in the absence of assistance (both questions that would require
an entirely different experimental design involving a large number of NGO-led
interventions).
Instead, the goal is to identify the presence of learning-by-exporting.
Second, given the nature of our experiment, we are unable to distinguish
exporting from selling to sophisticated domestic buyers—i.e.
domestic buyers who demand high quality and possess knowledge about how
to achieve such quality.
However, these buyers are scarce in developing countries, and as the literature
on quality upgrading we discuss below suggests, the presence of such buyers
may be the most pronounced difference between internal and external trade
for developing-country firms.
These two characteristics of exporting—the fact that not only is knowledge
transferred but there are also incentives in place to encourage firms to
absorb and implement this knowledge—explain why exporting to high income
markets can be a particularly effective way of transferring technology
to developing countries.

Our results relate to a number of papers that span the trade and development
literatures.
Most directly, we contribute to a voluminous literature that seeks to identify
the existence of learning-by-exporting.
The evidence from these studies is mixed
[^2]

, in part due to the severe selection and measurement issues highlighted
above.
We directly confront selection through random assignment; and directly
confront measurement both by collecting very detailed data on the production
process and by setting up a quality lab that allows us to perfectly control
for product specifications.
In doing so, we follow
SyversonJEL

and
BloomVanReenan2010

who advocate improving our understanding of productivity through more careful
measurement.

Our findings also relate to the literature on quality upgrading.
Studies using country- or product-level data show that product quality
positively co-varies with export destination income-per-capita (
Schott (2004)

,
Hallak (2006)

and
Hallak (2010)

); and firm-level studies suggest that exporting exposes firms in developing
countries to sophisticated buyers who demand higher quality.
[^3]

Unlike much of this literature that must infer quality from price data
or certifications, or through structural models that back out quality from
prices and quantities, we collect direct measures of quality.
[^4]

In addition, the combination of our RCT, the quality lab and the rich survey
data allow us to contribute to this literature by showing quality upgrading
occurs, at least in part, through improvements in technical efficiency
rather than through movements along the PPF alone.

Finally, although the use of RCTs is novel in the trade literature, the
methodology has been used to understand supply constraints in firms (e.g.


and
Bloom et al. (2013)

explore credit constraints, input market frictions and managerial constraints).
We complement this literature by providing the first experimental evidence
for the importance of demand constraints and the effects of relaxing those
constraints through expanding market access.

The rest of the paper is organized as follows.
Section

describes the research setting.
Section

explains our experimental intervention and introduces the data.
Section

examines the impact on profits and Section

decomposes the profit changes.
Section

presents a theoretical framework that then guides our five step approach
to detecting learning-by-exporting.
Section

concludes.

Research Setting
\label{S_setting}

Finding a Viable Setting
\label{ss_background}

In order to carry out a randomized evaluation of the impact of exporting,
we partnered with Aid to Artisans (ATA), a U.S.-based NGO with a mission
to create economic opportunities for producers of handmade products in
developing countries.
Discussions began in October 2009, and ATA had just received funding to
implement a market-access program in Egypt.

ATA's program in Egypt followed their standard protocol for generating successfu
l exporting relationships between small-scale developing-country producers
and high-income markets.
First, ATA explores the country in question for products that would both
appeal to consumers in high-income markets and be priced competitively.
Once candidate products are found, ATA identifies a lead intermediary based
in the developing country.
The lead intermediary assists in finding small-scale producers to manufacture
the products, is the conduit for passing information and orders between
the producers and the buyers, and handles the export logistics required
to ship the products to importers or retailers abroad.
ATA then works closely with the intermediary to design and market appealing
products.


Working through a lead intermediary firm, rather than matching individual
producers directly with foreign buyers, is an important aspect of ATA's
model.
By matching local intermediaries with foreign buyers, the intermediaries
can aggregate orders to spread the fixed costs of exporting across many
small producers.
The ultimate objective is to foster self-sustaining relationships whereby
ATA can pull out and the local intermediaries maintain or even expand their
export sales.

The process of exporting via an intermediary is common for handmade products
and among small firms more generally.


For example, Chinese customs data show that 52 percent of Chinese exports
in the specific HS code that the rugs in our study are classified under—HS
570231 (“Carpets and other textile floor coverings, wool”)—go through intermedi
aries (compared to 20 percent of overall Chinese exports).
[^5]

Across sectors, World Bank Enterprise Surveys (2006-2014) reveal that 62
percent of exporters with 5 or fewer employees (and 36 percent of all exporters
) use an intermediary to export.

Alongside ATA, we searched for viable Egyptian products and identified handmade
carpets from Fowa as having potential.
In terms of the industry choice, both the handmade craft industry and the
rug industry are large and important sources of employment in Egypt, as
well as in many other developing economies (see

).
Turning to the location, Fowa is a peri-urban town with a population of
65,000 located two hours southeast of Alexandria.


The town is well known for its carpet cluster which contains hundreds of
small firms that use wooden foot-treadle looms to manufacture flat-weave
rugs—a product in which Egypt has a strong historical reputation.


Crucially, we also identified a local firm, Hamis Carpets, to serve as
the lead intermediary.
Hamis is the largest intermediary in Fowa and accounts for around 20 percent
of the market.
At the time, Hamis earned 70 percent of its sales in the domestic market,
mostly selling to distributors and retailers in Cairo, Alexandria and Luxor.

The firms in Fowa typically consist of a single owner who operates out of
a rented space or sometimes his (all producers in our sample are men) home.
Family members or hired labor assist with setting up the loom and the finishing
stage.
The process of producing rugs is standardized across firms.
The two key inputs beyond the loom and labor are
*warp thread*
—wool or cotton thread that spans the length of the rug and is not visible
on the final rug but is necessary to hold the rug together and
*weft thread—*
the colorful threads weaved between the warp threads using a shuttle.


provides additional details regarding the production process.

Firms self-identify as specialists in one of four flat-weave rug types:
duble (the focus of this paper), tups, kasaees and goublan.
[^6]

The average duble rug destined for domestic markets is sold by firms for
LE42.5 (about $7 at the prevailing exchange rate)
[^7]

and requires 5.9 hours of labor, per m

.
After accounting for input costs, hourly wages are roughly LE3 ($0.48).

Within a particular rug type, quality can vary substantially.
There


are two determinants of quality.
First, higher quality is associated with more demanding specifications.
Specifications are codifiable attributes of the rug that are typically
chosen by the buyer; for example, the number of colors, the thread count
or the type of input thread.
Figure

shows one such specification sheet from a foreign buyer in our experiment.
Second, higher quality is associated with better weaving technique.
For example, how flat the rug lies on a hard surface is determined by how
skillfully the warp and weft thread are installed on the loom, and whether
the threads are held correctly while weaving.
Similarly, how well defined the corners are, how accurately the design
was followed, and whether the rug adheres to the desired size specifications
depend on weaving skill.
These attributes of the rug are both difficult to codify and depend on
the skill of the firm.

Generating Export Orders
\label{ss_hard}

It took the combination of ATA and Hamis more than two years to generate
sustained export orders from clients in high-income countries.


Generating sustained orders was not guaranteed.
The textile market is competitive and conversations with ATA's staff revealed
that only 1 in 7 matches lead beyond trial orders.
This is consistent with
(EatonTyboutSearch2014)

who estimate that only 1 in 5 importer-exporter matches result in successful
business relationships.

ATA first brought the CEO of Hamis to the US for a training course that
covered marketing and distribution.
At the same time, ATA hired an Italian consultant to design rug samples.
Hamis was then provided with marketing support—both by displaying these
samples


at various international gift fairs and by directly introducing Hamis to
foreign importers or retailers through a US-based rug intermediary.


Hamis and potential foreign buyers would discuss pricing, delivery time,
and product specifications (design, colors, materials, and so forth; recall
Figure

provides an example of how these specifications are codified after these
discussions).
Hamis would then organize the production of sample orders, either from
its in-house weavers or from one of the treatment firms in our sample.

The majority of rugs demanded by foreign buyers are duble rugs, although
one client ordered kasaees rugs.
There have been no orders for goublan rugs, even though the local market
in Egypt perceives these rugs to require the most skilled weaving techniques;
the painting-like style of goublan rugs is unlikely to appeal to buyers
in high-income countries (see Figure

for an example).
Likewise, domestic-style duble rugs did not attract interest from abroad.
Instead, it appears that foreign buyers


prefer
"

modern
"

designs, as illustrated in Figure

.

After one-and-a-half years of searching, in June 2012 Hamis Carpets secured
its first large export order (3,640m

) from a German buyer.
As of June 2014, multiple buyers


continue to place large, regular orders.
Figure

reports that cumulative export production between December 2010 and June
2014 (the end of our experiment) has totaled

[Table file not found: string".tex]

m

, resulting in cumulative payments to the producers of LE

[Table file not found: string".tex]


($

[Table file not found: string".tex]

).
[^8]


As described in the next section, these orders were entirely sourced from
our treatment firms, which forms the basis of our experiment.

The Experiment
\label{S_experiment}

Experimental Design
\label{ss_design}

We designed the following export-market-access intervention.
We drew a sample of small rug-producers (described in more detail in the
next section).
The firms were divided into two groups, treatment and control.
As described above, the local intermediary, Hamis Carpets, secured export
orders with ATA's marketing assistance.
Treatment firms were then visited by a representative of Hamis Carpets
and provided the opportunity to fill an
*initial*
export order.
More precisely, Hamis Carpets showed them the rug design, explained that
the carpet would be exported to a high-income country, and offered them
an order of 110m

which translates to about 11 weeks of work.
The 110m

was chosen by trading off the desire to have a reasonable sized initial
order and the need to have enough orders to treat the firms.
Hamis was free to choose the price offered to the producers based on the
specifications of the rugs (we analyze prices below).


To ensure all rug orders were consistent across producers, Hamis provided
the input thread and two loom components that are design-specific, the
reed and the heddle (

describe these components).
[^9]

At the same time, as is typical in buyer-producer relationships, Hamis
would discuss the technical aspects of the specific rug order and answer
any questions the firm may have.
Firms would deliver rugs to Hamis with payment upon delivery.

As further export orders were generated, we tried to mimic a normal buyer-seller
relationship as closely as possible.
Hamis was allowed to continue to place orders with treatment firms but
was not bound to make subsequent purchases from firms whose quality was
below par or who could not deliver on time.
(And firms were informed of this arrangement.)


In other words, the experiment protocol simply required Hamis to offer
an
*initial*
order to the treatment firms.


In contrast, the control firms were not contacted by Hamis Carpets about
an initial order, and Hamis Carpets was forbidden from sourcing from them
through the duration of the experiment.
[^10]

The experimental protocol is described in more detail in

.

As we show later, very few firms had ever knowingly exported at baseline.
Thus, the intervention provided treatment firms with the opportunity to
produce rugs for the export market.
Comparing outcomes between treatment and control firms allows us to identify
the causal effects of being provided with such an opportunity.
[^11]

An alternative experiment would be to provide our control firms with a similar
quantity of rug orders but from domestic rather than foreign sources.
We did not pursue this approach for reasons both theoretical and practical.
Trade models typically model exporting as a demand shock (e.g.

Krugman (1979)

,
MelitzEMA2003

), sometimes with features distinct from domestic demand shocks.
Increasing demand is also the primary motivation for many export promotion
policies (e.g.
sending trade delegations or analyzing opportunities for domestic firms
in foreign markets).
Therefore, to assess the impacts of exporting, it is natural to include
the increased demand it brings.
In terms of the practical limitations,


if we were to provide equally-sized initial domestic orders, it is unclear
on what dimension they should be equal given the different profit margins
and hours required per rug.
And then it would have been almost impossible to match subsequent foreign
and domestic orders over time, not least because of the fluctuations in
the Egyptian economy that would have made it extremely difficult to obtain
$

[Table file not found: string".tex]


of firm orders


from new domestic sources.

Sample Details and Takeup


\label{ss_takeup}

To obtain the sample of firms, we carried out a recruitment drive in Fowa
in July 2011.
To be eligible, the firm had to have fewer than 5 employees, work on their
own account (meaning that they bought their own inputs when an order required),
and never previously worked with Hamis.
The recruitment drive generated a sample 303 firms that specialized in
one of the four rug types described in Section

.
Anticipating that we would not secure orders for every type of rug, we
stratified the sample on the type of rug produced and the loom size (which
determines the maximum width of rug that can be produced).
Within each stratum we randomized firms into treatment and control.
For reasons that will become clear momentarily, we refer to these 303 firms
as
"

Sample 1
"

.
Table

show the total number of firms by rug type and treatment status.

It proved difficult to secure sufficient export orders to treat every firm
in the treatment group with an 110m

initial order.
As detailed in Section

, we were only successful in generating large and sustained orders for one
of the four rug types, duble rugs, and even then only one and a half years
after the baseline survey.


For this reason, we were only able to offer firms the opportunity to produce
duble rugs and we could only offer multiple smaller orders of 10-20m

spread out over many months instead of the full 110m

order in one go.
This resulted in very low takeup (16 percent) among specialized producers
of the non-duble rug types (firms not taking up cited an unwillingness
to switch rug types; see

).
Even among the 79 specialist duble producers, only 14 of the 39 treated
firms firms took up with non-takeup firms citing an unwillingness to jeopardize
their existing dealer relationships for the small 10-20m

order sizes.

Once large and sustained export orders for duble rugs arrived in June 2012,
it was feasible to offer the full 110m

initial order in one go.
Given the opportunity to implement the experimental design we originally
intended, we asked our surveyors to locate every remaining duble firm in
town that satisfied our sample criteria.
We found 140 additional duble firms in this manner which we refer to as

"

Sample 2
"

.
We stratified on loom size as before, and randomly selected 35 firms for
the treatment group (with the number of treated firms determined by Hamis'
capacity constraints).
Consistent with the claim that takeup in Sample 1 was low due to the small
order sizes, column 5 of Table

shows that 32 of the 35 treated Sample 2 firms took up when offered the
full initial order in one go.
All of the duble firms in both Sample 1 and 2 that took up the opportunity
were
"

successful
"

in the sense that they both delivered the 110m

that constituted the initial order and then received subsequent orders
from Hamis.

Given that we only secured large and sustained export orders for duble rugs
and that very few non-duble firms were willing to manufacture this rug
type, we were essentially unable to treat the non-duble strata and so exclude
them from our analysis.
To be clear, if the focus of this paper was to simply evaluate an export-market
-access program, it would be important to understand why the intervention
only generated sustained exports for one of the four products.
Instead, we use the random variation our experiment generates to investigate
a long-standing question in international trade: does exporting improve
firm productivity?

For the analysis in the rest of the paper, we combine the firms in the duble
strata


from Sample 1 with the firms in Sample 2—who are exclusively duble producers-–t
o form the
*Joint Sample*
of 219 firms (74 in treatment, 47 of which took up).
This is essentially all the duble producers in Fowa willing to identify
themselves to our surveyors (minus ten firms who refused to participate).
For completeness,

presents the results for duble firms in the two samples separately,
[^12]

as well as results for the non-duble strata.

Data
\label{ss_data}

We collected multiple rounds of data at around 4-month intervals.
In total, Sample 1 was interviewed seven times and Sample 2 four times.
The baseline round occurred before treatment firms were provided with the
opportunity to export and included questions on: (a) firm production; (b)
rug quality; and (c) household and demographic characteristics.
We repeated this full survey for follow-up round 5 of Sample 1 and follow-up
round 3 of Sample 2.
In all other follow-up rounds we administered an abbreviated survey on
firm production and rug quality.
The survey timeline is shown in Table

.

The production module records production activity for the month preceding
the survey interview.
All nominal variables are converted to real values using the official Egyptian
CPI.
We collect measures of profits, revenues, expenses, output quantity and
prices, input quantity and prices, total labor hours worked, and the specificat
ions of the rugs produced that month.
These specifications include: (1) the type of rug being produced; (2) how
difficult the rug is to make rated on a 1-5 scale by a master artisan accompany
ing the surveyors (see below); (3) the amount of weft thread used per m

of the rug (thread count); (4) the type of weft thread used (e.g.
Egyptian wool, cotton, etc.); (5) the number of colors used in the rug;
and (6) which segment of the market the rug is aimed at as reported by
the master artisan (normal, mid, or high).

The quality module records the quality of the rugs being produced by firms
at the time of the survey.
Rug quality is assessed by a master artisan under our employ who is a well-know
n and respected member of the rug community in Fowa.
Quality was measured along 11 dimensions: (1) corners; (2) waviness; (3)
weight; (4) touch; (5) packedness; (6) warp thread tightness; (7) firmness;
(8) design accuracy; (9) warp thread packedness; (10) inputs; and (11)
loom.
[^13]

The quality along each dimension is scored on a 1 to 5 scale, with higher
numbers denoting higher quality.
These quality metrics capture differences across rugs that are vertical
in nature; for example, at equal prices, both foreign and domestic consumers
would prefer a flatter-lying rug or a more accurate design.


As discussed in Section

, higher quality scores reflect a combination of better specifications and
greater weaving skill.

For takeup firms, a second quality module is available at higher frequency.
These firms deliver rugs to Hamis on a weekly basis.
Upon receiving each batch of rugs, Hamis records the size accuracy, design
accuracy, packedness, and weight accuracy.

We collected a third set of quality and productivity measures in June 2014
by asking firms to manufacture an identical domestic rug using identical
inputs and a common loom in a work space we had rented (our quality lab,
see Section

for further details).
Production was timed and upon completion the rugs were measured, anonymized
and then sent to be scored along the same quality dimensions


by both the master artisan and a Professor of Handicraft Science from Domietta
University.

Finally, we collect information on knowledge flows—the number of visits
Hamis made to firms and what was discussed during each visit—that we discuss
in Section

.

Summary Statistics
\label{ss_sumstats}

Table

shows baseline balance between the treatment and control groups.
The table reports regressions of each variable on a treatment dummy and
strata fixed effects, and reports the constant (the mean of the control
firms) and treatment coefficient (the difference between control and treatment
means).
Panel A shows summary statistics for the household characteristics of the
firm owner.
The mean age in the control group is around 51 years and, on average, firm
owners have slightly more than 37 years of experience working in the rug
industry.
Roughly 63 percent of firm owners are illiterate.
The average household size is 4.2 and the average total monthly household
income from all activities is LE1090 ($173).

Panel B reports statistics from the rug business.
Monthly profits from the rug business average LE646 ($102) in the control
group.
Firms report 247 labor hours in the previous month, which amounts to around
22 days of work at 11 hours per day.
As noted earlier, firm sizes are small because this was an explicit criterion
in choosing our sample: the average firm has just over one worker.
Total output per month is 50m

and only about 12 percent of firms have ever knowingly produced rugs for
the export market.
The final row of Panel B reports the average rug quality across the 11
dimensions.

Across both panels we find no statistical differences between treatment
and control firms with one exception: treatment firms report lower quality
scores at baseline.
The final row of Table

reports attrition across survey rounds (Table

reports attrition by round).
Attrition has been relatively low with a non-response rate of approximately
11 percent per round which does not vary across treatment and control groups,
while attrition in the quality lab is similar at 14 percent and balanced
across groups.

Causal Impacts of Export-Market Access on Profits
\label{S_profit}

Empirical Specifications
\label{ss_spec}

The randomization methodology allows us to use a straightforward specification
to assess the impact of the export-market access on firm profits:

$$
y_{it}=\alpha_{1}+\beta_{1}Treatment_{i}+\gamma_{1}y_{i0}+\delta_{s}+\tau_{t}+\varepsilon_{it},e_itt
$$


where
$y_{it}$

is the profit measure,
$Treatment_{i}$

is an indicator variable that takes the value 1 if firm
$i$

is in the treatment group,
$\tau_{t}$

are time period fixed effects,
$\delta_{s}$

are strata fixed effects and
$y_{i0}$

is the value of the dependent variable at baseline.
[^14]

We are essentially combining all follow-up survey rounds to increase precision,
and we cluster standard errors at the firm level to take account of the
fact that errors may be correlated within firms.
As

controls for the baseline value of the dependent variable, we cannot include
observations from the baseline survey in the regression.
[^15]

Since not all firms who were offered the opportunity to export took up
that offer,

is an intent-to-treat (ITT) specification.

We also present results from the treatment-on-the-treated specification
(TOT) which scales up the treatment effect to take account of the fact
that not everyone took up the treatment:
[^16]


$$
y_{it}=\alpha_{2}+\beta_{2}Takeup_{it}+\gamma_{2}y_{i0}+\delta_{s}+\tau_{t}+\nu_{it},e_tot
$$


where
$Takeup_{it}$

takes the value 1 if a firm took up the opportunity to export.
This is a time-varying measure that turns on when a firm first produces
carpets for the intermediary and stays on subsequently.
Of course takeup is not random and may be correlated with unobservables,
and so we instrument
$Takeup_{it}$

with the variable
$Treatment_{i}$

that is uncorrelated with the error (and the baseline control) thanks to
the randomization procedure.

Before showing results on profits and other metrics, we first show that
indeed the intervention worked, in so far as treatment firms were more
likely to manufacture rugs for export markets.
To do so, we replace
$y_{it}$

with a dummy variable that takes the value 1 if a firm ever knowingly made
rugs for export.
As shown in Table

, being in treatment raises the probability of ever exporting by 55 percentage
points from a baseline of 13 percent.


We also report the TOT specification, which suggests even more dramatic
increases.
[^17]

Profit Results


\label{ss_Measuring-Profits}

Following
de Mel et al. (2009)

—who assess the performance a variety of methods to elicit profits from
small firms—we construct four profit measures (described in more detail
in

).
The first measure directly asks firm owners to report profits from the
previous month (excluding wage payments to themselves).
This is the measure most strongly advocated by
de Mel et al. (2009)

as it avoids error due to the timing mismatch between revenues and expenses.
The second measure constructs profits from two survey questions that ask
firms to report their total revenues and total costs from the previous
month.
The third measure constructs profits from the production modules that contain
detailed information on prices and quantities of inputs and outputs.
The fourth measure, hypothetical profit, is based on a question that asks
firms how much profit they would earn from selling rugs if they purchased
a specific quantity of inputs.

The top panel of Table

shows the results of running specifications

and

on logged values of the four profit metrics, and for each we report the
ITT and TOT.
The first two columns (1A and 1B) report the specifications using the (log)
direct monthly profit measure.
The ITT coefficient is 0.26 and significant at the 1 percent level, implying
that the export treatment increases monthly profits by approximately 26
percent.
The TOT coefficient is, not surprisingly, larger at 42 percent and also
statistically significant.
Column 2 reports specifications using the profit measure constructed from
total revenues and costs in the previous month and Column 3 reports specificati
ons using the profit measure constructed from prices and quantities.
The ITTs are 21 and 19 percent, respectively, very similar to the direct
profits results.
This similarity suggests that any timing mismatch between revenues and
expenses is not severe in our setting where firms store little inventory,
and that rug inputs are not diverted to household consumption.
Finally, we examine the hypothetical profit measure in column 4.
Although these estimates are slightly higher (the ITT is 37 percent), it
is reassuring that all four measures increase by economically and statistically
significant amounts.
We also note that the treatment effects reflect profits rising more among
treatment firms rather than profits falling among control firms; control
firm profits increased in real terms across baseline and post-baseline
survey rounds (regressing log profits of control firms on a post-baseline
dummy yields a coefficient of

[Table file not found: string".tex]


(s.e.
of

[Table file not found: string".tex]

)).

These regressions indicate that the export treatment causally increases
measured profits by between 19-37 percent.
Of course, profits may have risen partly because firms increased their
labor hours.
This is an issue for our profits measures since most firms are owner-operated
and the profit measures do not account for the implicit wages paid to the
owner.
If we focus on profits per owner hour, these concerns are mitigated since
the value of the owners' time will be similar across treatment and control
due to the randomization.
[^18]

Therefore, the lower panel of Table

examines profits per owner hour by dividing each profit variable by the
total hours worked by the owner (or other unpaid family members when recorded)
in the previous month.


Using the direct profit per owner hour measure in columns 1A and 1B, we
find that the ITT estimate is 20 percent and again significant at the 1
percent level.
This estimate is lower than the corresponding estimates for profits which
implies that owners of treatment firms worked more hours.
The remaining columns also show lower estimates.
However, the basic message remains the same: the opportunity to export
raised profits per owner hour by 16-25 percent.
The differences between the panels suggests that total owner hours increased
by around 5 percent, a result we confirm in the next section when we analyze
labor hours.

Before turning to mechanisms, and in particular whether or not these improvement
s in firm performance occur through learning-by-exporting, we note that
it is not surprising that providing firms with a demand shock increases
profits.


What is surprising is the magnitude of the effect.
Many supply-side interventions on similar samples of firms have had limited
profit impacts.
A recent literature, surveyed by
McKenzie and Woodruff (2013)

, has carried out impact evaluations of business training programs for small
firms.
These programs had a statistically significant impact on profits in only
two out of nine studies that measured profits.
Another popular supply-side intervention is expanding access to credit.
The literature on the impacts of credit on profits for small firms also
finds mixed results.
For instance,
de Mel et al. (2008)

find returns to capital of around 5 percent per month while
Banerjee (2013)

cites several credit interventions that produced no statistically significant
increases in profits.
One possible interpretation of these mixed results and our findings of
substantial increases in profits is that supply-side interventions may
only be effective where there are no constraints on demand.
Thus, demand-side interventions such as our market-access program may be
complementary to more-standard supply-side interventions.

Sources of Profit Changes
\label{S_pi-sources}

Prices, Output, and Inputs


\label{ss_pi-sources}

This section explores the proximate sources of the increase in profits.
To fix ideas consider the following profit function for a firm:

$$
\max_{l,k}\pi=px(l,k)-wl-rk-Fe_pi-basic
$$


where
$p$

is the price a firm receives for one unit of rug.
The quantity of rugs produced is
$x,$


$w$

is the wage paid for each hour of labor
$l$

,
$r$

is the rental rate on capital
$k$

, and
$F$

is a fixed cost of production.
Although we analyze inputs below, we do not include input costs in

since a large majority of firms (91 percent


) receive raw material inputs from their intermediary and hence do not pay
for these expenses.
[^19]

Panel A of Table

uses our survey data to examine these various components of profits.
Columns 1A and 1B evaluate the impact of the intervention on the log output
price.
The ITT specification indicates 43 percent increase in prices with the
opportunity to export while the TOT indicates a 78 percent increase.
Thus, part of the profit increase from exporting is coming from significantly
higher prices per m

.
Columns 2A and 2B examine the impact of the opportunity to export on the
log total output weaved by the firm in the previous month (measured in
m

and unadjusted for product specifications).
The ITT estimate is -26 percent while the TOT is -47 percent; there is
a large
*decline*
in output in treatment firms relative to control firms.

Columns 3 to 5 document the impact of the intervention on firm scale, as
captured by labor and capital usage.
Column 3 shows the log of total hours
$l$

worked by all employees in the firm in the previous month.
The ITT estimate indicates a labor increase of 5 percent and the TOT is
8 percent.
This increase in labor hours comes on the intensive margin: as shown in
column 4 there is no change in the number of employees (inclusive of the
owner).
Since most firm owners are the primary weavers, and helpers are often family
members, we have very few observations of the wage
$w$

that may also be responding to the opportunity to export.
(We already showed that profits per owner hour increase but this combines
the shadow wage with firm profits).
In contrast to labor, we find no increase in capital usage
$k$


, as measured by the log number of active looms in use (column 5).
The fact that expansion occurs primarily along the intensive margin suggests
there may be large non-convexities associated with hiring additional workers,
particularly since an additional weaver is likely to need his own loom
and requires owners to manage a full-time employee for the first time.

Finally, we turn to fixed costs
$F$

in columns 6A and 6B.
We try to capture economies of scale via a proxy for the length of the
production run.


Firms place a warp thread ball on the loom at the beginning of a production
run.
A larger warp thread ball enables firms to amortize the costs of re-stringing
the loom over more units.
We find that the size of the warp thread ball increases by 15 percent in
the treatment group, indicating that the opportunity to export lowers fixed
costs


through longer production runs that require less frequent re-stringing.

Panel B of Table

examines input prices and quantities.
As noted above, most firms do not purchase the material inputs, but we
did ask these firms to estimate the price of the weft and warp thread inputs.
The first two columns use these data to explore the impact of the intervention
on thread prices.
Reported weft thread prices increase 20 percent.
In contrast, there is no evidence that warp thread prices are higher among
treatment firms.
These two findings are sensible given the production technology.
The warp thread is critical to maintain the rug structure but is not observable
in the finished rug.
Meanwhile, the weft thread is observable and can vary by both material
type (cotton, wool, polyester, etc.), material grade (e.g.
Egyptian wool or New Zealand wool


) and thickness.
Note that although columns 3-4 suggest that input quantities (measured
in grams) do not increase with the opportunity to export, the output decline
implies that the rugs produced by treatment firms use more material inputs
and are heavier than those produced by control firms.

The increases in prices, labor input usage and the length of production
runs appear consistent with two workhorse models used to study international
trade.
Comparative advantage models, such as the Ricardian model, would predict
that export prices are higher for products that Egypt has a comparative
advantage in (and it is reasonable to think handmade flat-weave rugs are
such a product).
In this framework, the opportunity to export would also raise the quantity
of labor being used in rug production, as we find.
Similarly, our findings on scale and fixed costs are consistent with a
standard scale effects story whereby exporting enables firms to reach larger
markets and spread fixed costs over more units (e.g.

Krugman (1979)

).
However, the
*reduction*
in output is not consistent with either of these frameworks.
The results are also not consistent with exporting simply being a generic
demand shock (which would also yield an increase in output).

The reductions in output accompanied by rising output (and input) prices
point to export-induced quality upgrading.
If high-quality rugs require more labor inputs, rug output can fall alongside
increasing revenues and input usage.
If high-quality rugs also require more expensive inputs (as shown by
Kugler and Verhoogen (2012)

), the rise in material input prices provides further evidence of quality-upgrad
ing.
In the next subsection we confirm this conjecture.

\textbf{
Quality and Unadjusted Productivity Measures
\label{ss_qlevels}

We first draw on the detailed quality metrics described in Section

to confirm that treatment firms are indeed manufacturing higher quality
products.
We have 11 different quality metrics that are ranked on a 1-5 scale with
5 being the best for that dimension of quality.

Table

presents the quality results.
Instead of implementing specifications

and

separately for each quality metric, we regress a stack of all 11 metrics
on interactions of the treatment (or takeup for the TOT) with indicator
variables for each quality metric.
We also include interactions of the quality-metric indicators with baseline
values, strata fixed effects and round fixed effects.
The resulting coefficients are identical to those from running separate
regressions for each quality metric, but run this way


we can cluster standard errors by firm to account for any firm-level correlatio
ns within quality metrics across time or across quality metrics within a
period.

For 10 of the 11 quality metrics, quality is significantly higher among
treatment firms (all at the 1 percent level).
The one exception is loom quality.
The lack of a treatment effect on loom quality is consistent with our understan
ding of the technology for rug production.
Although the loom size determines the maximum rug width, it matters little
for rug quality.

Since it is difficult to parse all 11 quality metrics separately, Panel
B of Table

restricts the coefficients on the treatment dummy to be identical across
all 11 quality metrics (recall they were all run in a single stacked regression
).
[^20]

Given the previous results, it is not surprising that we obtain positive
and statistically significant ITT and TOT estimates when we do this.
On average, quality (on a scale of 1 to 5) is 0.79 points higher among treatment
firms.
These are substantial increases in quality given a standard deviation of
quality of 0.55 at baseline.

We also examine two simple productivity measures, measures that do not adjust
for changes in product specifications or quality: unadjusted output per
labor hour and unadjusted total factor productivity (TFP).
Unadjusted output per labor hour comes from firms' responses to the question:

"

how long does it take you to make 1m

?
"

The second measure also accounts for capital inputs (although recall there
is limited variation in capital across firms with 92 percent of firms using
only one loom).
[^21]

Specifically, unadjusted TFP is equal to the residual from an estimated
Cobb-Douglas production function that includes both labor and capital (


see

for a full description of the procedure).
Table

shows that both of these productivity measures fall in treatment relative
to control.
Looking at the ITTs, unadjusted output per hour is 24 percent lower and
the unadjusted TFP is 28 percent lower, with even larger TOT effects.

Quality-Upgrading Mechanisms
\label{ss_mechanisms}

The finding that quality rises and unadjusted productivity falls alongside
rising profits is consistent with two different quality-upgrading mechanisms,
and the distinction is important for understanding how exporting improves
firm performance.
In the first mechanism, firms always knew how to manufacture the high-quality
rugs demanded by rich-country buyers.
If foreign buyers pay higher prices, but particularly so for high-quality
products, firms will upgrade quality as long as the returns offset any
costs (e.g.
more expensive inputs or more labor inputs).
This is a movement
*along*
the PPF.
Under this mechanism, the export opportunity raises the relative price
of high-quality rugs and profit-maximizing firms respond by producing rugs
with specifications associated with high-quality.
What does
*not*
change through this mechanism is technical efficiency.

While it is challenging to provide a direct mapping between profit margins
and quality levels, we provide some suggestive evidence for this phenomenon
by analyzing Hamis' (self-reported) cost structure for domestic and foreign
orders.
Hamis reports 9 percent profit margins on (lower-quality) domestic orders
and substantially higher margins of 33 percent on (higher-quality) foreign
orders.
(The full cost structure is broken down in Table

).
Although these are Hamis' profit margins, not the firms, if there is profit
sharing there would be a similar relationship between profit margins and
rug quality for producers.

A second mechanism is learning-by-exporting, which we follow the literature
and define as an export-induced change in technical efficiency (
Clerides et al. (1998)

,
Loecker (2007)

).
This is a shift
*out*
in the PPF and can include
*both*
transfers of information from buyers to producers and learning-by-doing
that would not have happened in the absence of exporting (e.g.
if export products have steeper learning curves).
If such changes in technical efficiency are biased towards the production
of high-quality rugs, quality upgrading can also occur through these learning
processes.

We emphasize that these two mechanisms
*are not*
mutually exclusive.
In fact, a rise in the price of quality is potentially a precondition for
the learning-by-exporting described above.
In these contexts, where the opportunity to export raises the price of
quality, learning-by-exporting generates further profit increases beyond
those from simply moving along the PPF.
In the next section, we define learning-by-exporting more precisely and
provide evidence it is present in our setting.

Detecting Learning-by-Exporting
\label{s_LBE}

A Framework for Detecting Learning-by-Exporting
\label{ss_model}

In order to be explicit about the learning-by-exporting mechanism, we enrich
the profit function as follows:

$$
\max_{l,k,\lambda}\pi & =px-wl-Fe_profits\\
x & =a(\lambda;\chi_{a})f(l,k)e_x\\
q & =q(\lambda;\chi_{q})e_q\\
p & =p_{0}+bq(\lambda)e_p
$$


Rug output
$x$

and rug quality
$q$

are determined by separate production functions.
Prices,
$p$

, are determined by a price function that is exogenous to the firm and is
increasing in the quality of the rug, with
$b>0$

determining the price of quality.

Each of the two production functions depend on efficiency parameters,
$\chi_{a}$

and
$\chi_{q}$

, that capture the skill of the firm, as well as on a choice variable: the
product specifications of the rug indexed by
$\lambda$

.
(Recall that specifications are codifiable rug attributes that buyers agree
upon before ordering; see Figure

for an example of such an agreement.) High-
$\lambda$

rugs have more demanding specifications, in the sense that they require
more inputs, primarily labor hours, to produce, and we assume that these
high-
$\lambda$

specifications are also associated with high-quality rugs.

More precisely, the production function for output
$x$

has two components.
Labor and capital inputs are mapped to output through
$f(l,k)$

and output per unit input is determined by the function
$a(\lambda;\chi_{a})$

, a TFP metric that is
"

unadjusted
"

for rug specifications.
[^22]

Unadjusted TFP
$a(\cdot)$

is decreasing in
$\lambda$

since rugs with more demanding specifications require more inputs and increasin
g in the efficiency parameter
$\chi_{a}$

which governs how quickly a firm produces rugs of a particular set of specifica
tions with a given set of inputs.
Collecting these two derivatives:

$$
\partial a(\lambda;\chi_{a)}{\partial\lambda}<0\qquad\partial a(\lambda;\chi_{a)}{\partial\chi_{a}}>0.e_assumpA-1
$$


Quality is determined by the function
$q(\lambda;\chi_{q})$

which we assume is increasing in product specifications as quality is achieved
in part through more demanding specifications.
Additionally, quality increases in the efficiency parameter
$\chi_{q}$

which governs a firm's ability to make quality given a particular set of
specifications.
Collecting these two derivatives:

$$
\partial q(\lambda;\chi_{q)}{\partial\lambda}>0\qquad\partial q(\lambda;\chi_{q)}{\partial\chi_{q}}>0.e_assumpQ-1
$$


With this structure in hand, it is straightforward to clarify what constitutes
a movement along the PPF due to exporting and what constitutes a shift
out (i.e.
learning-by-exporting).
Firms move along the PPF when there is an increase in
$b$

, the price of quality (due to, for example, foreign buyers higher willingness
to pay for quality).
This leads firms to choose higher specifications
$\lambda$

, and by

, quality rises.
In contrast, learning-by-exporting occurs when exporting raises
$\chi_{a}$

and/or
$\chi_{q}$

, the two efficiency parameters, and hence shifts out the PPF.
As mentioned earlier, this process can occur as firms move into high-quality
products with steep learning curves or through transfers of knowledge from
foreign buyers to domestic producers.
We expect transfers of knowledge about quality,
$\chi_{q}$

, to be particularly relevant for firms in low-income countries that sell
to buyers in high-income countries since these buyers are likely to demand
such high-quality products and possess knowledge about how to produce them.
Despite the different theoretical implications, we are unaware of earlier
work that seeks to distinguish these two quality-upgrading mechanisms.

To see that this theoretical framework can generate reductions in unadjusted
TFP alongside improvements in quality through either mechanism, we rearrange
the total derivatives of the first order conditions with respect to
$\lambda$

and
$l$

.
First, consider an increase in the price of quality
$b$

.
As long as there are diminishing marginal returns to raising specifications
(i.e, concavity of
$q$

and
$a$

in
$\lambda$

), firms choose to raise specifications and so equilibrium quality
$q^{*}$

rises and equilibrium unadjusted TFP
$a^{*}$

falls:
[^23]


$$
dq^{*}{db}=q_{\lambda}d\lambda{db}>0\qquadda^{*}{db}=a_{\lambda}d\lambda{db}<0.
$$


Now consider an increase in the quality efficiency parameter
$\chi_{q}$

.
As long as the complementarity between
$\lambda$

and
$\chi_{q}$

in producing quality is sufficiently large—i.e.
weaving skill is particularly valuable for the production of high specification
rugs,
$\partial^{2q}{\partial\lambda\partial\chi_{q}}>0$

(e.g.
higher thread-counts require more dexterity to weave)—firms also choose
to raise specifications and so
$q^{*}$

rises and
$a^{*}$

falls:
[^24]


$$
dq^{*}{d\chi_{q}}=q_{\chi_{q}}+q_{\lambda}d\lambda{\partial\chi_{q}}>0\qquadda^{*}{d\chi_{q}}=a_{\lambda}d\lambda{d\chi_{q}}<0.
$$


Intuitively, the strong complementarity ensures that it is profitable for
the firm to raise specifications since quality, and hence price per m

, increases faster than
$a$

declines.
The same pattern can arise with an increase in the output efficiency parameter,

$\chi_{a}$

, as long as there is a sufficiently large complementarity between
$\lambda$

and
$\chi_{a}$

in producing output—
$\partial^{2a}{\partial\lambda\partial\chi_{a}}>0$

; for example, skill matters more for speed when producing less-familiar
high-specification rugs—to ensure that the slowdown from raising specifications
exceeds the direct increase in speed from higher
$\chi_{a}$

.
[^25]

Our model does not allow for investments that raise the
$\chi$

parameters.


For the production function for physical output,
$x$

, investments that increase output


should be fully captured by increases in some type of capital or labor
in a well-specified production function.
If the returns to such investments rise with the opportunity to export,
any resulting changes should not be classified as learning-by-exporting
under our definition since these potential investments were already accounted
for in the PPF.
(A similar argument can be made for the quality production function if
we amend

to include labor and capital.)


Hence, purchasing a more efficient weaving machine or paying for a training
course in response to the export opportunity would not be considered learning-b
y-exporting.
In contrast, tacit knowledge passed on by a buyer or intermediary which
is neither anticipated nor paid for by the firm, even implicitly, would
be.
[^26]

Such a categorization is consistent with the learning-by-exporting literature
that considers these types of knowledge transfers archetypal.

Empirically detecting learning-by-exporting is challenging for two reasons.
First, firms with high efficiency parameters are likely to self-select
into export markets making it difficult to disentangle treatment effects
of exporting from selection (

).
The most convincing analyses to date rely on matching techniques which
requires that researchers fully specify the underlying selection model
(e.g.
see
Loecker (2007)

).
Here, we exploit the randomization to ensure that the opportunity to export
is uncorrelated with initial levels of
$\chi_{a}$

and
$\chi_{q}$

.

Second, even if self-selection were not an issue, researchers typically
measure technical efficiency through residual-based TFP.
TFP measures that do not adjust for prices (which is rarely the case) may
suggest learning-by-exporting when firms are just moving along the PPF
or obtain a higher markup in export markets.


In the few cases where price adjustments are made, measuring quantity-based
TFP requires comparing products with identical specifications and quality-level
s.
This is typically achieved by focusing on homogenous goods like concrete
and block ice (e.g.


) where trade frictions and learning may be limited; or by categorizing
products based on


administrative classifications and using a demand model to infer quality
from prices and market shares


(e.g.


).
In contrast, we exploit our rich panel data and quality lab to solve these
measurement issues.

We test several implications of the framework to detect learning-by-exporting:

- In Step 1, we use our detailed data on product specifications to show that
although our unadjusted productivity measure—corresponding to
$a(\cdot)$

above—falls with the opportunity to export (recall Table

), specification-adjusted productivity
*rises—*
consistent with
$\chi_{a}$

increasing.
We also show that specification-adjusted quality rises consistent with

$\chi_{q}$

increasing.
If there is no learning-by-exporting, specification-adjusted productivity
and quality should be
*unchanged*
as there is no change in the efficiency parameters.

- In Step 2, we demonstrate that when asked to produce identical-specification
domestic rugs using the same loom and the same inputs, treatment firms
produce higher quality products and do not take longer to do so.
Again, if there is no learning-by-exporting, treatment and control firms
should produce identical-specification domestic rugs at the same quality.

- In Step 3, we use time-series data to establish that quality and productivity
evolve over time as cumulative export production increases, consistent
with a learning process.
In contrast, if firms simply moved along the PPF we would expect a discontinuou
s jump upon exporting as firms immediately move to new quality and


productivity levels.

- In Step 4, we draw on correspondences between foreign buyers and Hamis,
as well as a log book of discussions between Hamis and the firms, to document
that our results come, in part, from knowledge transfers (information that
would be irrelevant if firms were only moving along the PPF).
In particular, we show that treatment firms improve quality most along
the particular quality dimensions that are discussed during meetings with
Hamis.
This evidence also strongly suggests that learning-by-exporting is not
solely driven by learning-by-doing (triggered by the export orders), but
in part through transfers of knowledge.

- In Step 5, we rule out alternative hypotheses that there were adjustment
costs, scale effects, or that firms made investments to raise output and
quality that were not adequately captured by our labor and capital measures.
In particular, we show that treatment firms make no monetary or time investment
s in upgrading, and do not pay, even implicitly, for the knowledge they
receive from the intermediary.

Step 1: Conditioning on Rug Specifications
\label{ss_step1}

If firms are only moving along the PPF, changes to unadjusted productivity
and quality should occur only through changes in rug specifications:
$da{db}|_{\lambda},dq{db}|_{\lambda}=0$

.
That is, producers always knew how to produce the particular rugs demanded
by foreign buyers, but previously chose not to because domestic buyers
did not value these rugs.
If there is learning-by-exporting,
\family roman
\series medium
}
}
\emph off
\bar no
\strikeout off
\uuline off
\uwave off
\noun off
then we would expect
\family default
}
}
}
}
\bar default
\strikeout default
\uuline default
\uwave default
\noun default
productivity and/or quality to rise, conditional on rug specifications,
due to an increase in
$\chi_{a}$

or
$\chi_{q}$

:
$da{d\chi_{a}}|_{\lambda},dq{d\chi_{q}}|_{\lambda}>0$

.

To separate these two hypotheses, we now repeat the quality and productivity
regressions above but control in various ways for the specifications of
the rug being manufactured at the time of the survey visit.
Recall from Section

that we have six dimensions of rug specifications.
Although imperfect, we note that many studies simply control for product
differences through product fixed effects based on statistical classifications.
Our first specification—the type of rug—is the analogous control, although
it uses a much finer classification than standard trade classifications
(e.g.
all of our seven rug types would fall within the U.S.
HS ten-digit classification 5702311000)


.
The remaining specifications, such as thread count or design difficulty,
are rarely observed by researchers.
Such controls are possible because there is overlap in rug specifications
across firms selling to domestic and foreign markets.
This overlap can be seen in Figure

which plots the distribution of each of the six specifications separately
for firms that are producing rugs for export (
\family roman
\series medium
}
}
\emph off
\bar no
\strikeout off
\uuline off
\uwave off
\noun off
i.e.

$Takeup_{it}=1$


\family default
}
}
}
}
\bar default
\strikeout default
\uuline default
\uwave default
\noun default
) and those that are not.
Note that if our specification controls are very crude, that will tend
to bias our findings towards the unconditional results we found in Tables


and

.
Hence, the prediction that productivity should rise conditional on specificatio
ns is particularly informative since unadjusted productivity falls.
(Step 2 deals more directly with the possibility that the specification
controls are imperfect.)

We present three sets of results.
The first approach regresses quality and the two unadjusted productivity
measures on treatment (or treatment instrumented with takeup) as in Section


but now also includes controls for the six specifications and their baseline
values.
The second approach goes further towards ensuring that the treatment and
control firms we compare are making identical rugs by including fixed effects
for each of the 435 unique combinations of the six specifications.
Since approximately one third of firm-round pairs are making unique rugs,
the cost of this approach is that we lose a significant number of observations.
The third approach follows from equation

which suggests that we can directly infer
$\chi_{a}$

from the residual of a production function estimation that includes specificati
ons.
As discussed in more detail in

, we use control firms to estimate this production function and then calculate
specification-adjusted TFP for each firm and round using the estimated
coefficients on labor, capital and specifications.
We construct similar measures for quality and output per hour by regressing
these variables on rug specifications in the control group and use the
resulting coefficients to construct adjusted metrics (actual minus predicted).
This procedure also partially addresses a second issue.
Although treatment is exogenous by design, the specification controls in
the first two approaches may be endogenous.
Of course, if higher-ability firms selected into higher-specification rugs
in the control group, the coefficients on rug specifications will be biased.
However, if anything, this bias will lead us to find no productivity gain.
[^27]

Panel A of Table

reports the approach with specification controls, Panel B the approach
with fixed effects for specification combinations, and Panel C the approach
with specification-adjusted dependent variables.
Before discussing the effects on quality and productivity, it is reassuring
to note that the specification controls in Panel A have the signs we assumed
in the model: More difficult rugs are associated with higher quality and
lower unadjusted productivity, while those destined for lower segments
of the market are associated with lower quality and higher unadjusted productiv
ity.
And, the


R-squared rises substantially compared to the regressions without specification
controls in Tables

and


(increasing from 0.39 to 0.64 in the quality ITT, and from 0.18 and 0.26 to
0.57 and 0.62 in the productivity ITTs


);


this suggests


that the rug specifications have substantial explanatory power.

Turning to the treatment effects in Table

, recall that without conditioning on specifications (i.e.
Panel B of Table

and Panel A of Table

) quality rises and the productivity measures fall.
Conditioning on specifications using the three approaches described above,
quality again rises significantly but the signs on the productivity measures
flip from significantly negative to significantly positive in all three
cases.
(The productivity ITTs correspond to productivity increases between 14
and 31 percent.) That is, conditional on making similar rugs, treatment
firms are making them faster than control firms.
These results suggest a rise in the efficiency parameters
$\chi_{a}$

and
$\chi_{q}$

.

Step 2: Production of Identical-Specification Domestic Rugs (the Quality
Lab)
\label{ss_step2_lab}

The second step exploits our experimental setting to compare quality and
productivity across firms producing identical-specification domestic rugs
(rather than relying on specification data to control for the type of rug).
If firms are only moving along the PPF, when asked to make rugs with identical
specifications, quality and productivity should not differ across treatment
and control firms (since treatment was randomly assigned).
In order to carry out this test we brought the owners of each firm to a
rented workshop in June 2014 and asked them to produce an identical-specificati
on rug using identical inputs and the same loom.


We chose rug specifications that mimicked a popular rug design sold at
mid-tier domestic retail outlets in Egypt (the rug is shown in Figure

) and told the firms the orders were for a new buyer in Cairo


.
The master artisan assigned a difficulty rating of 3 for this rug (below
the 4.28 average rating of export orders).
We hired a new staff member to implement the quality lab


and gave identical instructions to treatment and control firms.
Owners were given a fixed payment a little above the market wage to compensate
for producing in an external location.
We provide the experimental protocol used for this step in

.

As discussed in Section

, each completed rug was given an anonymous identification number and the
master artisan was asked to score it along the 9 of the 11 quality dimensions
discussed previously (we provided the inputs and the loom so the last two
dimensions were not relevant).
The identification system ensured that the master artisan had no way of
knowing whether the rug was made by a treatment or control firm.


We also sent the rugs to be scored by a second external quality assessor,
a professor of handicraft science at Domietta University located two hours
from Fowa, to cross check the accuracy of the master artisan's scoring.

Panel A of Table

reports ITT and TOT results separately for each of these 9 quality metrics.
[^28]

Using the master artisan's scores, quality is significantly higher among
treatment firms for all dimensions.
Reassuringly, treatment firms also score significantly higher along every
dimension using the professor's quality assessments.

Panel B of Table

constrains the coefficients on treatment to be the same across all the
quality metrics.
Unsurprisingly, this average effect is positive and statistically significant.
In terms of magnitudes, the point estimate from the master artisan's scores
is 0.64 and the standard deviation is 0.75 among the control group, implying
that the opportunity to export increases quality levels by 0.85 standard
deviations.


For the professor's scores, the increase is 0.5 standard deviations.

Panel C of Table

reports the accuracy of rugs in terms of the length, width and weight that
we requested.
We define these variables as the negative of the absolute deviation from
the target value, so higher values reflect greater accuracy.
Treatment firms produce rugs that are closer to the requested length and
weight.
We do not observe statistical differences in the width of the rugs, but
this is expected since the loom size determines the width (and all firms
used the same loom).

Finally, we recorded the time taken to produce the rug.
Since the rug specification, material inputs and loom are identical for
all firms in this setup, the time taken reflects firm productivity.
The 4th row of Panel C shows that, on average, firms took 4 hours to produce
the rug.
Although the ITT is not significant, treatment firms took 6 minutes less.
That is, despite manufacturing rugs with higher quality metrics, treatment
firms spend if anything less time weaving, not more.

In the absence of learning-by-exporting, we would not expect differences
between treatment and control firms when producing identical-specification
rugs for the domestic market using the same inputs, the same loom, and
at the same scale.
If anything we might expect control firms to produce these rugs quicker
or at higher quality since they have recent experience manufacturing domestic
designs and specifications.
It also seems unlikely that treatment firms put more effort into weaving
the rug because they were worried poor performance would jeopardize their
relationship with Hamis.
Firms were not informed of any link between the quality lab and Hamis Carpets–-
recall we hired a new staff member to run the lab in order to disassociate
it from the export opportunity randomization as much as possible and we
told firms that the order was from a new buyer in Cairo—and if firms did
believe there was a link it is just as plausible that control firms put
in extra effort to impress Hamis in order to gain export orders.

In contrast, we find strong evidence of higher quality levels among treatment
firms that persist even when manufacturing rugs for the domestic market,
indicative of an increase in
$\chi_{q}$

.
As treatment firms do not take longer to produce these rugs, these results
imply that a broader productivity measure that adjusts for both specifications
and quality would rise substantially.
Note that the lack of a significant treatment effect for time taken does
not contradict the increase in productivity we found in Step 1.
In the presence of a complementarity between
$\lambda$

and
$\chi_{a}$

, the additional skill acquired through exporting will only translate into
faster production for more demanding export rugs and not these simpler
domestic rugs that they were already familiar with.

Step 3: Learning Curves
\label{ss_learning}

The third step examines the time paths of quality upgrading.
Unlike a movement along the PPF which should be instantaneous (see Section


for a discussion of adjustment costs), learning processes typically take
time.


In the most obvious formulation, the efficiency parameters change with the
opportunity to export through the cumulative production of export rugs.
This captures the idea that efficiency improves with repeated interactions
with buyers and/or because learning curves are steeper among export rugs
that are less familiar to our sample firms.
Therefore, if there is learning-by-exporting, productivity and quality
is likely to rise with cumulative exports.
If there is no learning, although quality may immediately jump with the
first export order (or unadjusted productivity may fall), the levels should
remain constant with subsequent export orders.

To investigate potential learning curves in a non-parametric manner, we
carry out a two-stage procedure.
In the first stage, we regress our quality or productivity measures on
both firm and round fixed effects.
[^29]

In the second stage, we plot a kernel-weighted local polynomial regression
of the residuals against cumulative export production.
Since cumulative export production is only available for takeup firms,
we only include these firms in the second stage (although all firms are
included in the first stage when we demean by survey round).


presents similar plots using the partially linear panel data estimator
proposed by
Baltagi and Li (2002)

.

Figure

shows these residual plots for the productivity measures as well as the
stacked quality measure.
[^30]

The upper and middle left plots show the two unadjusted productivity measures;
the plots indicate a decline in productivity until about 600m

after which it starts to rise.
The second column uses the specification-adjusted measures described in
Section

.
Both specification-adjusted output per hour and TFP rise with cumulative
exports, consistent with the initial dip in unadjusted productivity being
driven by the move to more difficult product specifications demanded by
foreign buyers.
The bottom row of Figure

presents the analogous learning curves for the stacked quality measures.
Both plots show a


rise in quality up until 200m

of exports and then a leveling off.
The typical firm weaves 10-15m

per week, which suggests that firms learn how to produce the quality demanded
by foreigners within about five months.
Overall, the learning curves suggest faster learning about quality efficiency

$\chi_{q}$

than about output efficiency
$\chi_{a}$

.
[^31]

In contrast to the ITT results in Steps 1 and 2, the fact we allowed Hamis
to allocate follow-up orders means that more orders may have been given
to firms whose quality or productivity was improving.
With that caveat in mind, we see the evolution of productivity and quality
among takeup firms as suggestive of a learning explanation.
Figures

and

complement these results by showing similar patterns using four higher-frequenc
y quality metrics collected by the intermediary for each batch of rugs delivered
by each takeup firm.

Step 4: Knowledge Transfers
\label{ss_step4-infoflows}

The results in Steps 1-3 indicate that learning-by-exporting is present
in our context.
In this step, we distinguish between two types of learning-by-exporting
discussed in the literature.
The first is a learning-by-doing story where learning curves are particularly
steep for the high-quality items demanded by foreigners and so the learning-by-
doing is induced by exporting.
The second is a story where actual knowledge is transferred between buyers,
the intermediary and producers.
Of course, we believe both are occurring, and this subsection simply provides
evidence that some of the learning comes from knowledge transfers.

To measure knowledge flows, we tracked information flowing between buyers
and Hamis, as well as between Hamis and the producing firms.
The data on flows between buyers and Hamis come from email correspondences
Hamis shared with us.
Here we provide several excerpts relating to various aspects of rug quality.
In one correspondence, a foreign buyer complained that the rug was packed
too tightly which results in wavy rugs:

Wrapping the kelims tightly and strongly leaves waving marks on them, so
please roll kelims and wrap them softly to avoid waviness.

On a separate occasion, the same buyer also noted that the edges of some
carpets had frayed:

We have a problem with our client.
As you remember, this client asked for two carpets with fringes in the
colour uni 2 and 3.
Now after one and a half year using the carpets, the fringes crumble away,
as you see on the pictures [reproduced in Figure

].


They will have two new pieces and give the whole problem to an lawyer.
What to do?

(sic)

These conversations suggest that buyers are passing along both information
on how to manufacture high-quality rugs (e.g.
packing that is not too tight) as well as information on what a high-quality
product is (e.g.
the importance of long-term durability).
However, the correspondences are more suggestive in nature and are incomplete–-
we have no records of the information flows resulting from Hamis Carpets'
export experiences prior to our study (although Hamis reports having learned
about both weaving techniques and quality control from previous interactions
with foreign buyers).
[^32]

We have more detailed data on information flows between the intermediary
and the firms.
Hamis provided us with a log book of the visits made to each of the treatment
firms as well as the subject discussed during that visit.
In particular, we know the total number of conversations, their average
length, and the topics discussed over the project period (Table

presents summary statistics from this dataset).
The topics are categorized according to 10 of our 11 quality metrics (the
intermediary did not discuss input quality since it provided the inputs).


All takeup firms were visited at least seven times, with the average firm
visited 11 times.
A visit lasted 28 minutes on average.
They talked about issues related to design accuracy, the weight of the
rug and the tightness of the warp thread on at least half of the occasions.

To find out more about the nature of these discussions, in August 2014 we
asked firms whether their conversations along each of these dimensions
involved the intermediary providing
"

information on techniques to improve quality
"

or just involved the intermediary
"

pointing out flaws
"

.
The data show that 89.4 percent of discussions were of the former type,
discussions about production techniques to improve quality.
The intermediary has provided us with multiple examples of the techniques
passed on in these conversations.
For example, the intermediary provided knowledge about the optimal way
to weave the weft thread through the warp so as to achieve the correct
firmness of the rug, about how to hold the weft thread to reduce waviness,
and about how to maintain the integrity of the rug corners.

We can use these data to examine whether the knowledge imparted on these
visits correlates with the improved performance of takeup firms.
We match the dataset of topics discussed during visits with each firm to
the quality metrics recorded in the final survey round.
This match allows us to test whether quality increased most along the particula
r quality dimensions discussed with Hamis.
To perform this test, we once more stack the quality (or specification-adjusted
quality) measures, indexed by
$d$

, and run the following cross-sectional regression:


$$
Quality_{id=\alpha_{3}+\beta_{3}Takeup_{i}\times1[Talked_About_Dimension]_{id}+\gamma_{3}Quality_{id0}+\delta_{i}+\delta_{d}+\varepsilon_{id}.e_flows
$$


We include firm fixed effects
$\delta_{i}$

so that we are comparing improvements in quality across the different dimension
s
$d$

within the same firm.
We also include quality metric fixed effects
$\delta_{d}$

to control for different means across dimensions.
[^33]

A significant
$\beta_{3}$

coefficient is supportive of the presence of knowledge transfers as it
implies that quality improves more along dimensions that are discussed
than along dimensions that are not (and inconsistent with a simple movement
along the PPF, where quality would be independent of knowledge flows).

The results in Table

support the hypothesis that knowledge is transferred from the intermediary
to the firm.
We find a positive and statistically significant association between changes
in quality and whether the intermediary discussed that quality dimension
with the firm, using either our standard quality metrics in column 1 or
our specification-adjusted ones in column 2.
The size of the coefficient implies that quality levels improve by 16 percent
more when the intermediary spoke to the firm about that dimension.
[^34]

Columns 3 and 4 allow the
$\beta_{3}$

coefficient to differ depending on whether the intermediary provided informatio
n on techniques or was just pointing out flaws.
We find that both types of information are associated with improvements
in quality, with the magnitudes of the two coefficient virtually identical.
Finally, columns 5-8 repeat the analysis but include additional interactions
between
$Takeup_{i}$

and each quality metric to allow for differential treatment effects by
dimension.


The coefficients in columns 5-6 are almost identical to those in columns
1-2 and highly significant.
While the coefficients actually become larger than uninteracted coefficient
when the type-of-information interactions are included in columns 7-8,
the higher standard errors mean that significance levels decline, with
the technique interaction for specification-adjusted quality falling just
below the 10 percent threshold.

We provide one additional piece of evidence that suggests our results are
not driven by learning-by-doing alone.
Under learning-by-doing, we would expect firms who were already producing
high-quality rugs at baseline to see smaller treatment effects as they
had less to learn.
This prediction is not borne out by the data: When we regress the stacked
quality metrics on a treatment dummy, baseline quality and an interaction
of the two, the interaction coefficient is insignificant.
[^35]

It is hard to completely dismiss the possibility that these discussions
communicate what firms can get away with or the rug preferences of foreigners.
However, the picture painted by these knowledge flows results is fully
consistent with the quality lab results in Step 2.
The examples above of superior weaving techniques discussed during these
visits are all techniques that once acquired are costless to implement.
Given that the majority of discussions are of this type, it is little surprise
that treatment firms produced substantially higher quality domestic rugs
in the quality lab despite having no incentive to do so.

Step 5: Ruling Out Alternative Hypotheses
\label{ss_althypo}

In this final step, we rule out alternative explanations that could explain
the patterns in the data.


There are three main competing hypotheses.
The first is that firms incur an adjustment cost while moving along the
PPF which could generate learning curves of the type we found in Step 3.
While a reasonable story, adjustment costs alone cannot explain our findings
in the other steps.
For example, adjustment costs would not lead treatment firms to produce
identical domestic rugs at higher quality many months after the intervention
(Step 2).

The second hypothesis is that we are simply picking up scale effects poorly
captured by our production function.
But we found no evidence that, relative to control firms, treatment firms
increased scale by hiring more workers or expanding the number of looms
(see Table

).
Although treatment firms did increase scale through more labor hours and
longer production runs, it is implausible to attribute the Step 2 quality
lab results to this greater scale since every firm had the same-length
production run in the quality lab.
Moreover, using the estimated relationship between quality and hours worked
in the control group, the 5 percent increase in labor hours (from Table


) would increase quality by less than one hundredth of the difference between
treatment and control in the quality lab (from Table

).
[^36]

A third closely-related hypothesis is that the opportunity to export raised
the returns to investments that raise output or quality but are not adequately
captured by increases in our simple measures of capital and labor inputs.
These investments could take the form of purchasing equipment, investing
time in learning new techniques, or hiring consultants to teach new skills.
If we do not account for these investments, we may spuriously conclude
that there was learning-by-exporting.

Our data allow us to dismiss a simple investments hypothesis.
First, we regularly surveyed firms about investments or costs incurred
throughout the study.
There is no quantitative (or qualitative) support indicating that treatment
firms undertook any such investments.
For example, no firm reports investing in a new loom or paying to repair
existing looms over the duration of the sample.
Additionally, we asked treatment firms about the extent to which they practiced
weaving techniques, and none report ever practicing techniques.

A more complicated variant of the investments hypothesis would be that our
intermediary provided a teacher or consultant to train treatment firms
in weaving skills.
If the intermediary deducted training costs from payments to the firm this
would be equivalent to an investment by the firm.
However, we find no evidence of this type of payment: the price paid to
firms is uncorrelated with the number of hours the firm was visited by
the intermediary.
[^37]

Instead, the knowledge transfers occurring during these interactions appear
to be


flows of information that are not priced, exactly the type of information
flows described in the classic learning-by-exporting literature (e.g.

Clerides et al. (1998)

).

A final variant of the investments hypothesis is that firms invest their
time to raise the quality of their labor (i.e.
trade off slower production now for higher returns from exporting once
they can produce high-quality rugs with their improved labor).
Inconsistent with this hypothesis is the fact that the benefits of the
knowledge transfer, even in the domestic market, exceed the likely time
costs incurred if higher quality could be purposefully learned.
To determine the value of improving quality for a firm that sells to the
domestic market, we regress profits per hour for non-takeup firms (i.e.
those selling to the domestic market) on our quality metrics.
[^38]

Combining these estimates of the domestic returns to quality with the treatment
effects estimated in the quality lab, the benefit for control firms to
move to the quality levels achieved by treatment firms would be

[Table file not found: string".tex]


percent higher profits on the domestic market.


We assume that it takes 5 months to learn how to produce high quality (the
approximate time taken to weave the 200m

after which learning stops in the bottom row of Figure

).
Even if firms we assume that firms do not benefit at all from higher profits
per hour until this learning period is over, profits would fall by only
5 percent in the learning period (taking the productivity drop between
0 and 200m

in the first row of Figure

) and the firm would recover the investment in quality upgrading after 8
months


using an annualized discount rate of 10 percent


.
Also note that, using the wage the intermediary pays their employees who
visit the firms and assuming all the time spent on firm visits is spent
discussing techniques,


the cost of providing this training is only LE103 ($16) compared to a lifetime
net present value for firms of LE10,070 ($1,596).


Hence, these calculations suggest that if firms were choosing whether or
not to invest in learning the skills needed for high-skill production,
they should have already chosen to make this investment to produce for
the domestic market.

Conclusion
\label{sec:Conclusion}

This paper conducts the first RCT that generates exogenous variation in
the opportunity to export in order to understand the impacts of exporting
on firm performance.
The random variation, the detailed survey collection, and our quality lab
allow us to make causal inferences about the impact of exporting and to
identify the mechanisms through which improvements occur.

We find that profits for treatment firms increase 16-26 percent relative
to control.
This finding stands in contrast to many RCTs designed to alleviate supply-side
constraints that have shown limited impacts on profits.
Thus, our profit results suggest that demand-side constraints may be a
critical barrier to firm growth in developing countries and can be mitigated
through market access initiatives.
The question of whether this market access program is cost effective and/or
alleviates market failures is an interesting one which we leave for future
work.

The rise in profits is driven by substantial quality upgrading accompanied
by declines in output per hour.
However, this quality upgrading we observe may or may not come about through
learning-by-exporting—export induced improvements in technical efficiency.

We provide five pieces of evidence that learning-by-exporting is occurring
in our context.
First, conditional on product specifications, we observe large improvements
in both quality and productivity.
Second, when asked to produce an identical domestic rug, treatment firms
produce higher quality rugs and do not take longer to do so.
Third, we observe learning curves among the firms who took up the opportunity
to export.
Fourth, we document knowledge flowing between foreign buyers, the intermediary,
and the producers, with quality increasing most along the specific dimensions
that the knowledge pertained to.
Fifth, we find no evidence that firms make monetary or time investments
in upgrading, or pay, even implicitly, for the knowledge they receive.

Taken together, the evidence indicates that learning-by-exporting is present
in our data and that the learning occurs, at least in part, through information
flows.
Given that this learning is induced by demand for high-quality products
from knowledgeable buyers in high-income countries, these changes would
likely not have occurred as a result of increased market access to domestic
markets.

As is the case in any analysis of a particular industry or location, we
are cautious to generalize our findings too broadly.
However, we believe that two features of this study—random assignment of
export status and detailed surveys that allow us to unpack the changes
occurring within the firms—contribute to the literature that studies the
impacts of trade on the developing world.

\series medium
}


\start_of_appendix
\renewcommand Appendix \Alph{section}


Measuring Unadjusted and Specification-Adjusted TFP
\label{s_append-tfp}

 \theequation{A. equation}

A key challenge in standard productivity analysis is the lack of firm-specific
input and output prices which biases productivity estimates (see de Loecker
and Goldberg, 2014).
We avoid these measurement issues because we observe output quantities.
Moreover, since all firms produce a single product—handmade rugs—issues
that arise with multi-product firms and the divisibility of inputs are
not relevant in this setting (de Loecker et al., 2016).

The first production function estimation does not control for rug specifications
and hence provides our
*unadjusted TFP*
estimates.
We estimate the following Cobb-Douglas production function:


$$
\ln x_{it}=\lambda+\alpha_{l}\ln l_{it}+\alpha_{k}\ln k_{it}+a_{it}^{unadjusted}+\epsilon_{it}e_tfp_unadjust
$$


where
$x_{it}$

is the output (in m

) of firm
$i$

in period
$t$

,
$l_{it}$

is total labor hours used,
$k_{it}$

is the number of active looms, and
$a_{it}^{unadjusted}$

is the firm's unadjusted TFP.
The error term captures unanticipated shocks as well as an omitted variable,
the specifications of the rugs produced.

The second production function estimation controls for rug specifications
and provides our
*specification-adjusted TFP*
estimate.
We estimate

$$
\ln x_{it}=\lambda+\gamma_{l}\ln l_{it}+\gamma_{k}\ln k_{it}+Z_{it}^{'}\Gamma+a_{it}^{adjusted}+\nu_{it}e_tfp_adjust
$$


where
$a_{it}^{adjusted}$

is the firm's adjusted TFP and the vector
$Z_{it}$

includes six rug specifications: rug difficulty, thread count, thread type,
number of colors, market segment, and narrow product type.
[^39]

The error term now only captures unanticipated shocks (and, of course,
measurement error).

We do not include the costs incurred to find foreign buyers in the production
function for physical output.
Common to the learning-by-exporting literature, we view any productivity
gains from exporting as unanticipated by the firm.
That is, firms make exporting decisions without anticipating that they
would become more efficient producers.
The evidence in Section

supports this claim by showing that productivity improvements come, at
least in part, from knowledge transfers that the firms do not pay for.
In this case, the costs incurred to find foreign buyers


should not be included in the production function for physical output under
a reasonable definition of TFP.


[^40]

Although data on quantities and specifications deal with measurement concerns,
there is still potential simultaneity bias since TFP is observed by the
firm but not us.
We follow the control function approach (Olley and Pakes, 1996) and assume
capital is subject to adjustment costs, labor is a flexible input, and
we use warp thread quantity as the proxy.
We exploit the experimental design by estimating the production function
using only control firms, avoiding the need for assumptions on how treatment
changes the TFP process.
[^41]

We estimate the production functions using the one-step approach proposed
by Wooldridge (2009), with
$l_{it-1}$

as the instrument for
$l_{it}$

, and cluster standard errors by firm.


For production function (

), we obtain
$\alpha_{l}=$


[Table file not found: string".tex]


(s.e.
of

[Table file not found: string".tex]

) and
$\alpha_{k}=$


[Table file not found: string".tex]


(s.e.
of

[Table file not found: string".tex]

).
[^42]

For production function (

), we obtain
$\alpha_{l}=$


[Table file not found: string".tex]


(s.e.
of

[Table file not found: string".tex]

) and
$\alpha_{k}=$


[Table file not found: string".tex]


(s.e.
of

[Table file not found: string".tex]

).
We cannot reject the null of constant returns to scale in either case.

Having estimated the coefficients, we compute
$a_{it}^{unadjusted}=\ln x_{it}-\alpha_{l}l_{it}-\alpha_{k}k_{it}$

and
$a_{it}^{adjusted}=\ln x_{it}-\gamma_{l}l_{it}-\gamma_{k}k_{it}-Z_{it}^{'}\Gamma$

, where hats denote estimated parameters.
Note that the methodology assumes that the parameters of the production
function are the same for treatment and control firms.
This is a reasonable assumption because all firms produce rugs using the
same technology in all periods.
Since all firms produce a narrowly defined product, the assumption is weaker
than existing work that typically assumes identical parameters for all
firms within a 2- or 4-digit industry classification.


### Appendix References


\hangindent= \parindent \hangafter=1


\shape smallcaps
de Loecker, J.

}
(2013).

"

Do exports generate higher productivity? Evidence from Slovenia
"

,
*Journal of International Economics*
, 73, 69-98.


\hangindent= \parindent \hangafter=1


\shape smallcaps
de Loecker, J.

}
(2013).

"

Detecting Learning by Exporting
"

,
*American Economic Journal: Microeconomics*
, 5(3), 1-21.

\hangindent= \parindent \hangafter=1


\shape smallcaps
de Loecker, J.
and P.
Goldberg
}
(2014).

"

Firm Performance in a Global Market
"

,
*The Annual Review of Economics*
, 6, 201-227.

\hangindent= \parindent \hangafter=1


\shape smallcaps
de Loecker, J., P.
Goldberg, A.
K.
Khandelwal, N.
Pavcnik
}
(2016).

"

Prices, Markups and Trade Reform
"

,
*Econometrica*
, 84(2), 445-510.

\hangindent= \parindent \hangafter=1


\shape smallcaps
Olley, G.
S.
and A.
Pakes
}
(1996), “The Dynamics of Productivity in the Telecommunications Equipment
Industry,”
*Econometrica*
, 64, 1263–97.

\hangindent= \parindent \hangafter=1


\shape smallcaps
Woolridge, J.
M.

}
(2009), “On Estimating Firm-level Production Functions Using Proxy Variables
to Control for Unobservables,”
*Economics Letters*
, 104, 112–114.


**Table: t_takeup**


    Duble	Orders Kasaees	Orders
    Joint	Sample Sample	1 Sample	2 Sample	1
    Duble	Firms Goublan	Firms Tups	Firms Duble	Firms Duble	Firms Kasaees	Firms
    Statistic (1) (2) (3) (4) (5) (6)
    Firms 219 103 83 79 140 38
    Treatment	firms 74 49 42 39 35 19
    Takeup	firms 46 5 8 14 32 5
    Initial	packet	size	(m2) 110 110 110 110 110 250
    Successful	takeup	firms 46 4 6 14 32 5
    Mean	output	conditional	on	takeup	(m2) 538 586 589 778 434 303
    S.D.	of	output	conditional	on	takeup 188 174 208 132 177 76
    Notes: Table reports statistics by firm type and sample. The 1st row displays the number of firms within each rug type and sample. The 2nd row displays the
    number of firms in the treatment group. The 3rd row indicates the number of firms who accepted the treatment and agreed to make rugs for export. The 4th
    row is the initial order size (in square meters) offered to each takeup firm. The 5th row shows the number of firms that completed the initial order successfully
    and received subsequent orders from Hamis. The 6th row indicates average output conditional on takeup. The last row reports the standard deviation of
    output	in	the	first	year,	conditional	on	takeup.


**Table: \label{t_timeline**


    Survey Timeline Sample 1 Sample 2
    Baseline Round 0 *July-Aug 2011 *Feb-Mar 2013
    Follow Up Round 1 §Nov-Dec 2011 May-June 2013
    Follow Up Round 2 April-May 2012 Nov-Dec 2013
    Follow Up Round 3 Sept-Dec 2012 *May-June 2014
    Follow Up Round 4 Mar-Apr 2013
    Follow Up Round 5 *July-Oct 2013
    Follow Up Round 6 Jan-Mar 2014
    Quality Lab June 2014 June 2014
    Notes: Table reports the timeline for the data collection on duble firms by sample.
    *Supplementary questions about household and firm outcomes are included in both samples Baseline Round 0 surveys and Follow Up Round 5
    for Sample 1 and Follow Up Round 3 for Sample 2.
    §The Egyptian survey company we hired for the Sample 1 baseline trained a new enumerator for the first Sample 1 follow-up on. Unfortunately,
    we discovered that this enumerator did not actually interview all of the firms and entered in fake data for some, so we discard the data from
    Round 1 for Sample 1 . We immediately fired the enumerator and hired new employees in January 2012 who conducted all subsequent surveys.
    We managed these employees directly and implemented a stricter auditing procedure as well as back checks using external evaluators to ensure
    data integrity.


**Table: \label{t_balance**


    Control	Group Difference	in
    Mean Treatment N
    Panel	A:	Household	Characteristics
    Age 51.0 0.9 218
    (0.7) (1.6)
    Number	of	years	in	rug	business 37.7 0.2 213
    (0.8) (1.7)
    Illiterate? 0.63 0.10 214
    (0.03) (0.07)
    Household	size 4 .2 0.0 219
    (0.1) (0.2)
    Household	income 1 ,090.0 76.5 219
    (91.2) (228.0)
    Digit	Span	Recall 5 .8 0.2 204
    (0.1) (0.2)
    Panel	B:	Firm	Characteristics
    Price	per	square	meter 30.2 6.8 218
    (3.3) (7.8)
    Direct	monthly	profits	from	rug	business 6 46 7.9 218
    (41.8) (81.5)
    Reported	monthly	profits	from	rug	business 8 06 -10.4 217
    (38.5) (84.4)
    Hours	worked	last	month 2 47 -1.7 218
    (5.6) (11.7)
    Number	of	employees 1.09 0.0 218
    (0.0) (0.1)
    Total	produced	last	month	(m2) 50.0 3.3 218
    (4.3) (10.0)
    Ever	exported? 0.12 0.02 219
    (0.02) (0.05)
    Average	Quality 2.63 -0.13 *** 218
    (0.03) (0.05)
    Joint	F-test 1.23
    Attrition	in	Follow	Up	Surveys 0.11 0.00 815
    (0.01) (0.02)
    Attrition	in	Quality	Lab 0.14 0.02 219
    (0.03) (0.05)
    Notes: Table explores baseline balance. Each row is a regression of the named variable on a constant, treatment dummy and strata fixed
    effects; the constant (control group mean) and treatment dummy are reported. The 3rd to last row reports the F-test for a regression of
    the treatment dummy on all 14 baseline balance variables. Profits and prices are winsorized at the 1st and 99th percentile to trim outliers
    (without winsorizing, the sample still remains statistically balanced between treatment and control groups). The final rows report average
    attrition	rates	across	all	survey	rounds	and	in	the	quality	lab	respectively.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_firststage**


    ITT TOT
    (1) (2)
    Indicator	for	Ever	Exported 0.55 *** 0.76 ***
    (0.06) (0.07)
    R-squared 0.33 0.45
    Control	Mean 0.20 0.20
    Observations 191 191
    Notes: Table regresses an indicator for if a firm has ever knowingly produced rugs for export markets on indicators for treatment (column 1)
    or takeup (column 2). The question was asked in Round 5 for Sample 1 and Round 3 for Sample 2. The TOT regression instruments takeup
    with treatment. The regressions control for baseline values of the dependent variable, and include round and strata fixed effects. Significance
    *	.10;	**	.05;	***	.01.


**Table: \label{t_profits**


    Panel	A:	Profits		(in	Month	Prior	to	Survey)
    Log	(Reported	Revenues	- Log	(Constructed	Revenues	-
    Log	Direct	Profits Reported	Costs) Constructed	Costs) Log	Hypothetical	Profits
    ITT TOT ITT TOT ITT TOT ITT TOT
    (1A) (1B) (2A) (2B) (3A) (3B) (4A) (4B)
    Treatment 0.26 *** 0.42 *** 0.21 *** 0.37 *** 0.19 *** 0.34 *** 0.37 *** 0.68 ***
    (.05) (.08) (.06) (.10) (.06) (.10) (.11) (.19)
    R-squared 0.21 0.22 0.16 0.18 0.16 0.18 0.19 0.19
    Control	Mean	(In	Levels) 929 929 931 931 951 951 541 541
    Observations 573 573 644 644 685 685 687 687
    Panel	B:	Profits	per	Owner	Hour	(in	Month	Prior	to	Survey)
    Log	(Reported	Revenues	- Log	(Constructed	Revenues	-
    Log	Direct	Profits Reported	Costs) Constructed	Costs) Log	Hypothetical	Profits
    ITT TOT ITT TOT ITT TOT ITT TOT
    (1A) (1B) (2A) (2B) (3A) (3B) (4A) (4B)
    Treatment 0.20 *** 0.32 *** 0.17 *** 0.29 *** 0.16 *** 0.28 *** 0.25 *** 0.46 ***
    (.05) (.08) (.05) (.09) (.05) (.09) (.07) (.12)
    R-squared 0.14 0.14 0.12 0.13 0.13 0.13 0.19 0.18
    Control	Mean	(In	Levels) 3.53 3.53 3.54 3.54 3.55 3.55 5.56 5.56
    Observations 573 573 637 637 684 684 687 687
    Notes: Table reports treatment effects on different measures of real profits in the month prior to the date of the survey, all measured in logs. See text for
    descriptions of each measure. Dependent variable in Panel A is profits. Dependent variable in Panel B is profits per owner hour. Owner hours include the hours
    of family member production when recorded. The regressions control for baseline values of the dependent variable, and include round and strata fixed effects.
    Control group means are reported in levels in Egyptian pounds (LE) in Panel A and LE/hour in Panel B. The TOT regressions instrument takeup with treatment.
    Standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_components**


    Panel	A:	Components	of	Profits
    Log	Output	Price	(LE/m2) Log	Output	(m2) Log	Hours	Worked
    ITT TOT ITT TOT ITT TOT
    (1A) (1B) (2A) (2B) (3A) (3B)
    Treatment 0.43 *** 0.78 *** -0.26 *** -0.47 *** 0.05 ** 0.08 **
    (0.10) (0.19) (0.09) (0.17) (0.02) (0.04)
    R-squared 0.16 0.15 0.24 0.22 0.12 0.13
    Control	Mean	(In	Levels) 28.2 28.2 64.1 64.1 269.0 269.0
    Observations 691 691 676 676 678 678
    Number	of	Employees Log	Number	of	Looms Log	Warp	Thread	Ball	(kg)
    ITT TOT ITT TOT ITT TOT
    (4A) (4B) (5A) (5B) (6A) (6B)
    Treatment 0.01 0.01 -0.02 -0.04 0.15 *** 0.25 ***
    (0.01) (0.01) (0.04) (0.06) (0.05) (0.08)
    R-squared 0.02 0.02 0.13 0.13 0.24 0.24
    Control	Mean	(In	Levels) 1.0 1.0 1.1 1.1 6.0 6.0
    Observations 695 695 694 694 600 600
    Panel	B:	Inputs
    Log	Weft	Thread Log	Warp	Thread Log	Weft	Thread Log	Warp	Thread
    Price	(LE/m2) Price	(LE/m2) Quantity	(g) Quantity	(g)
    ITT TOT ITT TOT ITT TOT ITT TOT
    (1A) (1B) (2A) (2B) (3A) (3B) (4A) (4B)
    Treatment 0.20 *** 0.33 *** -0.04 -0.07 -0.19 ** -0.34 ** 0.04 0.06
    (0.06) (0.10) (0.03) (0.06) (0.10) (0.17) (0.11) (0.20)
    R-squared 0.22 0.24 0.27 0.27 0.23 0.22 0.10 0.11
    Control	Mean	(In	Levels) 12.8 12.8 18.1 18.1 110.0 110.0 17.8 17.8
    Observations 564 564 685 685 677 677 686 686
    Notes: Panel A reports treatment effects on output prices and quantities, hours, number of employees (inclusive of owner), looms and the size of the warp
    thread ball (which is a proxy for the length of the production run), all measured in logs except number of employees. Panel B analyzes input prices and
    quantities, all measured in logs. The TOT regressions instrument takeup with treatment. Hours worked are calculated using average daily hours and number of
    days worked last month. Control group means are reported in levels. The regressions control for baseline values of the dependent variable, and include round
    and	strata	fixed	effects.	Standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.
    Notes: Panel A reports treatment effects on output prices and quantities, hours, number of employees (inclusive of owner), looms and the size of the warp
    thread ball (which is a proxy for the length of the production run), all measured in logs except number of employees. Panel B analyzes input prices and
    quantities, all measured in logs. The TOT regressions instrument takeup with treatment. Hours worked are calculated using average daily hours and number of
    days worked last month. Control group means are reported in levels. The regressions control for baseline values of the dependent variable, and include round
    and	strata	fixed	effects.	Standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_quality**


    Panel	A:	Quality	Metrics
    Control ITT TOT
    Mean (1) (2)
    Corners 2.98 1.11 *** 1.70 ***
    (0.12) (0.11)
    Waviness 2.99 1.10 *** 1.68 ***
    (0.12) (0.10)
    Weight 3.08 1.07 *** 1.63 ***
    (0.11) (0.11)
    Touch 3.12 0.40 *** 0.66 ***
    (0.06) (0.07)
    Packedness 3.11 0.89 *** 1.59 ***
    (0.11) (0.12)
    Warp	Thread	Tightness 3.05 0.83 *** 1.49 ***
    (0.10) (0.12)
    Firmness 2.98 0.87 *** 1.60 ***
    (0.11) (0.12)
    Design	Accuracy 3.17 0.79 *** 1.41 ***
    (0.10) (0.12)
    Warp	Thread	Packedness 3.05 1.07 *** 1.65 ***
    (0.11) (0.11)
    Inputs 3.07 0.89 *** 1.62 ***
    (0.10) (0.12)
    Loom 2.02 0.03 0.05
    (0.02) (0.04)
    R-squared 0.44 0.60
    Observations 6,885 6,885
    Panel	B:	Stacked	Quality	Metrics
    Control ITT TOT
    Mean (1) (2)
    Stacked	Quality	Metrics 2.96 0.79 *** 1.35 ***
    (0.09) (0.08)
    R-squared 0.39 0.54
    Observations 6,885 6,885
    Notes: Panel A stacks the quality metrics and interacts treatment (ITT) or takeup (TOT) with a quality-metric indicator variable. The coefficients on the
    interactions provide the treatment effects seperately for each metric. The TOT instruments takeup interacted with quality metric with treatment interacted
    with quality metric. Each regression includes baseline values of the quality metric, strata and round fixed effects, and each of these controls interacted with
    quality-metric. Panel B constrains the treatment effects to be equal across quality metrics; these regressions include baseline values, strata and round fixed
    effects.	Control	group	means	are	reported	in	levels.	Standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_outputhr**


    Log	Unadjusted	Output	Per	Hour Log	Unadjusted	TFP
    ITT TOT ITT TOT
    (1) (2) (3) (4)
    Treatment -0.24 *** -0.42 *** -0.28 *** -0.50 ***
    (0.09) (0.16) (0.09) (0.16)
    R-squared 0.18 0.16 0.26 0.24
    Control	Mean	(In	Levels) 0.26 0.26 0.49 0.49
    Observations 687 687 674 674
    Notes: Table reports treatment effects for the two productivity measures: log unadjusted output per labor hour (in m2/hour) and log unadjusted TFP. See text and
    Appendix A for the methodology used to obtain unadjusted TFP. The TOT specifications instrument takeup with treatment. Control group means are reported in
    levels.	Regressions	control	for	baseline	values	of	the	variable,	round	and	strata	fixed	effects.	Standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_step1**


    Panel	A:	Specification	Controls
    Stacked	Quality	Metrics Log	Output	per	Hour Log	TFP
    ITT TOT ITT TOT ITT TOT
    (1) (2) (3) (4) (5) (6)
    Treatment 0.32 *** 0.78 *** 0.18 ** 0.44 ** 0.14 ** 0.35 **
    (0.04) (0.08) (0.08) (0.18) (0.07) (0.16)
    (log)	Thread	quantity 0.04 0.02 -0.12 -0.13 -0.07 -0.08
    (0.05) (0.04) (0.13) (0.13) (0.13) (0.12)
    Difficulty	Control 0.47 *** 0.34 *** -0.14 *** -0.21 *** -0.16 *** -0.22 ***
    (0.02) (0.03) (0.04) (0.05) (0.04) (0.05)
    (log)	#	colors 0.03 ** 0.01 -0.05 * -0.07 ** -0.06 ** -0.07 ***
    (0.01) (0.01) (0.03) (0.03) (0.03) (0.02)
    Low-market	Segment -0.19 *** -0.08 ** 0.43 *** 0.49 *** 0.42 *** 0.47 ***
    (0.03) (0.03) (0.08) (0.09) (0.07) (0.08)
    Mid-Market	Segment -0.19 *** -0.05 0.29 *** 0.36 *** 0.26 *** 0.32 ***
    (0.04) (0.04) (0.08) (0.09) (0.07) (0.08)
    Rug	Type	FEs yes yes yes yes yes yes
    Input	Thread	Type	FEs yes yes yes yes yes yes
    R-squared 0.64 0.67 0.57 0.58 0.62 0.63
    Observations 6,820 6,820 673 673 660 660
    Panel	B:	Specification	Fixed	Effects
    Stacked	Quality	Metrics Log	Output	per	Hour Log	TFP
    (1) (2) (3) (4) (5) (6)
    Treatment 0.13 ** 0.53 *** 0.31 *** 1.26 *** 0.25 *** 0.94 ***
    (0.05) 0.17 (0.08) (0.45) (0.08) (0.36)
    Specification	FEs yes yes yes yes yes yes
    R-squared 0.71 0.72 0.69 0.62 0.70 0.68
    Observations 6,820 6,820 428 428 416 416
    Panel	C:	Specification-Adjusted	Dependent	Variables
    Stacked	Quality	Metrics Log	Output	per	Hour Log	TFP
    (1) (2) (3) (4) (5) (6)
    Treatment 0.42 *** 0.72 *** 0.18 ** 0.33 *** 0.20 *** 0.36 ***
    (0.05) (0.04) (0.07) (0.13) (0.07) (0.12)
    R-squared 0.18 0.27 0.06 0.10 0.13 0.18
    Observations 6,860 6,860 678 678 669 669
    Notes: Table reports treatment effects for the stacked quality measures and the two productivity measures after including various controls for the
    specifications of the rug on the loom at the time of the survey. The TOT specifications instrument takeup with treatment. In Panel A, there are both rug and
    input thread type fixed effects in addition to the specification controls included in the table. Panel B uses fixed effects for each of the 435 combinations of
    the 6 specification controls. Both panels use unadjusted productivity measures. Panel C uses specification-adjusted meausre for quality and productivity;
    see text and Appendix A for details. Regressions in all panels also control for baseline values of the dependent variable (and baselince values of the
    specification	controls	in	Panel	A),	as	well	as	round	and	strata	fixed	effects.	Standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_lab**


    Panel	A:	Quality	Metrics
    Master	Artisan Professor
    Control ITT TOT Control ITT TOT
    Mean (1A) (1B) Mean (2A) (2B)
    Corners 3.23 0.72 *** 1.05 *** 3.31 0.29 ** 0.43 **
    (0.14) (0.17) (0.13) (0.18)
    Waviness 3.17 0.55 *** 0.80 *** 3.31 0.25 ** 0.36 **
    (0.14) (0.18) (0.12) (0.16)
    Weight 3.60 0.62 *** 0.91 *** 3.64 0.58 *** 0.86 ***
    (0.13) (0.16) (0.17) (0.25)
    Packedness 3.30 0.77 *** 1.14 *** 3.28 0.28 ** 0.42 ***
    (0.13) (0.15) (0.11) (0.15)
    Touch 3.29 0.52 *** 0.76 *** 3.27 0.36 *** 0.52 ***
    (0.11) (0.14) (0.12) (0.16)
    Warp	Thread	Tightness 3.00 0.51 *** 0.74 *** 3.30 0.25 ** 0.36 **
    (0.09) (0.11) (0.12) (0.16)
    Firmness 3.21 0.71 *** 1.04 *** 3.23 0.29 ** 0.43 ***
    (0.14) (0.17) (0.12) (0.16)
    Design	Accuracy 3.65 0.53 *** 0.77 *** 3.45 0.27 ** 0.40 ***
    (0.11) (0.15) (0.11) (0.15)
    Warp	Thread	Packedness 3.05 0.87 *** 1.28 *** 3.20 0.39 *** 0.58 ***
    (0.14) (0.17) (0.12) (0.16)
    R-squared 0.21 0.34 0.11 0.14
    Observations 1,680 1,680 1,667 1,667
    Panel	B:	Stacked	Quality	Metrics
    Master	Artisan Professor
    Control ITT TOT Control ITT TOT
    Mean (1A) (1B) Mean (2A) (2B)
    Stacked	Quality	Metric 3.28 0.64 *** 0.94 *** 3.33 0.33 *** 0.48 ***
    (0.10) (0.12) (0.10) (0.13)
    R-squared 0.19 0.32 0.09 0.13
    Observations 1,680 1,680 1,667 1,667
    Panel	C:	Objective	Metrics
    Control ITT TOT
    Mean (1A) (1B)
    Length	Accuracy -4.51 1.43 *** 2.09 ***
    (0.51) (0.71)
    Width	Accuracy -2.29 0.17 0.25
    (0.29) (0.41)
    Weight	Accuracy -221.0 89.1 *** 131.0 ***
    (20.3) (29.6)
    Time	(in	minutes) 247.0 -5.67 -8.3
    (6.6) (9.5)
    R-squared 0.84 0.84
    Observations 748 748
    Notes: Table reports ITT and TOT specifications using the 9 quality metrics from the quality lab. Panel A stacks the quality metrics and interacts treatment (ITT) or
    takeup (TOT) with a quality-metric indicator variable. The coefficients on the interactions provide the treatment effects seperately for each metric. The TOT
    instruments takeup (interacted with quality metric) with treatment (also interacted with quality metric). Panel B constrains the treatment effects to be equal
    across quality metrics. Columns 1A and 1B report scores from the master artisan. Columns 2A and 2B report scores from the Professor of Handicraft Science at
    Domietta University. Panel C reports objective accuracy measures which are calculated as the negative of the absolute error for that specification, so that a
    higher value indicates that the manufactured rug was closer to intended length (140 cm), width (70 cm) and weight (1750 g). It also includes the time spent to
    produce the rug in minutes. As in Panel A, these are run in a single regression by intereacting the objective measure with treatment or takeup. All regressions
    include	interactions	of	strata	fixed	effects	with	quality-metric	indicators,	and	standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_visits**


    Panel	A:	Baseline
    Specification- Specification-
    Stacked	Quality Adjusted	Quality Stacked	Quality Adjusted	Quality
    Metrics Metrics Metrics Metrics
    (1) (2) (3) (4)
    Takeup i x	{Talked	About	Dimension id 0.19 ** 0.16 **
    (0.08) (0.07)
    Takeup i x	{Information	on	Techniques	for	Dimension} id 0.32 *** 0.33 ***
    (0.09) (0.09)
    Takeup i x	{Pointed	Out	Mistakes	in	Dimension} id 0.30 *** 0.29 ***
    (0.04) (0.04)
    Quality	Metric	FEs yes yes yes yes
    Takeup i x	Quality	Metric	FEs no no no no
    Firm	FEs yes yes yes yes
    Specification-adjusted	Quality	Metrics no yes no yes
    R-squared 0.76 0.43 0.75 0.42
    Observations 1700 1667 1670 1637
    Panel	B:	With	"Takeup	x	Quality"	FEs
    Specification- Specification-
    Stacked	Quality Adjusted	Quality Stacked	Quality Adjusted	Quality
    Metrics Metrics Metrics Metrics
    (5) (6) (7) (8)
    Takeup i x	{Talked	About	Dimension} id 0.15 *** 0.13 **
    (0.05) (0.05)
    Takeup i x	{Information	on	Techniques	for	Dimension} id 0.16 * 0.16
    (0.09) (0.10)
    Takeup i x	{Pointed	Out	Mistakes	in	Dimension} id 0.17 ** 0.15 *
    (0.08) (0.08)
    Quality	Metric	FEs yes yes yes yes
    Takeup i x	Quality	Metric	FEs yes yes yes yes
    Firm	FEs yes yes yes yes
    Specification-adjusted	Quality	Metrics no yes no yes
    R-squared 0.81 0.56 0.81 0.55
    Observations 1700 1667 1670 1637
    Notes: Table regresses stacked quality metrics on takeup indicator and its interaction with a dummy that takes the value 1 if the intermediary talked to the firm
    about that particular dimension of quality. Even-numbered columns use the specification-adjusted quality metrics described in the text. Columns 3 and 4
    seperate whether the discussion along that dimension was about technique or just pointing out mistakes. Regressions are run on a cross-section of firms and
    include baseline values, firm fixed effects, quality metric fixed effects, and quality metric fixed effects interacted with takeup. Panel B repeats the regressions in
    Panel	A	but	also	includes	takeup	interacted	with	quality	metric	fixed	effects.	Standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Figure: f_specification**


**Figure: \label{f_domestic-export**


**Figure: \label{f_magdyorders**


\begin{sidewaysfigure[H]

Overlap in Rug Specifications between Domestic and Export Orders}


sidewaysfigure[t]

Learning Curves for Takeup Firms (Step 3)}


 \thefigure{B. figure}

 \thetable{B. table}


\renewcommand Online Appendix \Alph{section}


\size largest
Exporting and Firm Performance:

\textbf{
\size largest
Evidence from a Randomized Experiment

\size large
David Atkin, Amit K.
Khandelwal, and Adam Osman


\textbf{
\size largest
Appendices for Online Publication

}


The Global and Egyptian Rug Industry
\label{s_append-industry}

The handmade craft industry and the carpet and rug industry are large and
important sources of employment in developing economies.
Global handmade craft production was estimated at $23.2 billion in 2005,
while world production of carpets and rugs totaled $32 billion in 2008
(UNCTAD Creative Economy Report, 2010).
Egypt is the 11th largest global producer of carpets and rugs with a total
production of $734 million in 2013


(36 percent of Egypt’s total textile sector and 1.3 percent of total manufacturi
ng output).
[^43]

More than 17,000 people work in the carpet and rug industry in Egypt, represent
ing 7 percent of world employment in this industry and 1.7 percent of total
manufacturing employment in Egypt (UNIDO 2013).


Egypt has a revealed comparative advantage in this sector, and in 2013
Egypt's exports in HS57 (
"

carpets and other textile floor coverings
"

) constituted 1.4 percent of its total exports.
The top 5 destinations for these exports are the U.S., Germany, Italy, United
Kingdom, and Canada, which together account for 59 percent of total exports
in the category.
Indirect exporting is common in both the sector


and among small firms.
As noted in the main text, 2006-2014 World Bank Enterprise Surveys reveal
that 36 percent of exporters across countries and sectors use an intermediary
(and 62 percent for exporting firms with 5 or fewer employees).


Moreover, using Chinese customs data, we are able to obtain the share of
exports that go through intermediaries for the specific HS code that the
rugs in our study are classified under, HS 570231 (“Carpets and other textile
floor coverings, wool”).
[^44]

Exporting via intermediaries is particularly common in the rug industry
with 52 percent of exports in HS Code 570231 going through intermediaries
compared to 20 percent of overall exports.

The Rug Production and Marketing Process
\label{s_append-production}

 \thefigure{C. figure}

 \thetable{C. table}

The firms in Fowa typically operate out of a rented space or sometimes the
home of the owner.
The owner is almost always the primary weaver and most firms have no other
full time employees, although family members often assist with restringing
the loom and the finishing stage.
Firms self-identify as specialists in one of four flat-weave rug types:
duble, tups, kasaees and goublan.
Duble and tups rugs are the most common types, with Tups rugs typically
having more elaborate patterns than Duble rugs; kasaees rugs are woven
from rags and are the cheapest type; goublan rugs are the most expensive
type and are works of art used as wall hangings.
See Figure

for pictures of each type.

The process of producing rugs is standardized across firms.
The elements of the production technology are marked in Figure

.
The rugs are made on a large wooden foot-treadle loom.
The width of the loom determines the maximum width of a rug.
Rugs can be made of any length.
The
*warp thread*
is the wool or cotton thread that spans the entire length of the rug and
must be attached to the loom before rugs can be weaved as part of the restringi
ng process.
These threads cannot be seen on the final rug but are necessary to hold
the rug together.
The warp threads are kept in place using a
*reed*
which resembles a very large comb.
The
*weft thread*
(typically made from wool) is the visible thread on the rug and is weaved
between these warp threads using a shuttle.
A foot-operated
*heddle*
is used to raise every alternate warp thread allowing the weaver to more
quickly weave the weft threads between the warp threads.
The weaver changes out the weft thread as he weaves based on the needs
of the design until the rug is complete.
At that point he cuts off the completed rug and continues to utilize the
remaining warp thread until the production run of that particular type
of rug is finished.
Hence, the length of the production run is determined by the amount of
warp thread in the warp thread ball attached to the loom.

Duble rugs are the thickest of the rug types and typically made using a
reed with 200 openings per meter.
In contrast, tups and goublan rugs are more intricate.
Goublan rugs usually require a reed with 400 openings per meter and because
of the intricate design patterns, weavers use their hands, instead of shuttles,
for precision in weaving the weft thread.
Tups rugs are also finely woven using a reed with 400 openings per meter
but because they are used as floor rugs and have simpler designs, weavers
use shuttles to increase the speed of production.
Finally, kasaees rugs use reeds with 250 openings per meter but these rugs
use left-over cloth (for example, torn up t-shirts) in their weave instead
of more expensive wool inputs.

The vast majority of rug producers in Fowa use intermediaries to market
their products.
[^45]

The baseline survey revealed that 88 percent of firms sell products through
one of the intermediaries located in Fowa.
We find no statistical difference in the proportion of firms using intermediari
es between treatment and control, nor is there a statistically significant
difference in the use of intermediaries by firms who had exported at baseline
and those who had not.
Thus, the reliance on intermediaries is not unique to exports.
(And as noted earlier in

, the use of intermediaries more generally is common, particularly for small
firms).
Finally, the baseline data indicate that the (limited number of) firms
who report having manufactured rugs for export markets have lower unadjusted
output per hour and higher quality metrics, just as we found for firms
induced to export by our intervention.
[^46]

This evidence suggests that, for the firms in Fowa at least, the process
of exporting through Hamis Carpets is similar to to the process of exporting
through other means.


**Figure: f_rugpics**


**Figure: \label{f_technology**


\section{Additional Details on Experiment Protocol
\label{s_append-details}

In this appendix, we provide a detailed description of the experiment protocol.
For completeness, we repeat the details reported in the main text, as well
as provide additional information.

Experiment Protocol
\label{subsec:Experiment-Protocol}

As described in the main text, Hamis Carpets marketed rugs to overseas buyers
with ATA's assistance.
[^47]

Once export orders were secured, we divided the orders into smaller packets.
The treatment firms were visited by our survey team and a representative
of Hamis carpets, and were offered the opportunity to produce one of these
packets of export orders.
More precisely, Hamis Carpets showed them the rug design, explained that
the carpet would be exported to a high-income country, and offered them
an order of 110m

which translates to about 11 weeks of work.
The 110m

was chosen by trading off the desire to have a reasonable sized initial
order and the need to have enough orders to treat all the treatment firms.
Hamis was given discretion regarding what price to offer the treatment
firms, and chose prices based on the specifications of the particular rug
orders.
The initial protocol intended for us to offer this 110m

packet in one go, and for the rug orders to be of the type the firm was
specialized in producing.
As we discuss in

, this was the protocol followed for Sample 2.
Given our initial difficulties in generating export orders, in Sample 1
we could only offer duble and kasaees rugs orders, and only by sequentially
offering smaller packet sizes that summed to 110m

.
[^48]

If the firm accepted, Hamis delivered the input thread and the correctly
sized reed and heddle to ensure all rug orders were consistent across producers.
At the same time, as is typical in many buyer-producer relationships, Hamis
would discuss the technical aspects of the specific rug order and answer
any questions the firm may have.
Firms would deliver rugs to Hamis and receive payment upon delivery.

As further export orders were generated, Hamis continued to place them with
the treatment firms.
Just as in any arms-length transaction, after the initial order amounts
were offered, Hamis was not bound to continue to make subsequent purchases
from any particular treatment firm if the quality was below par or the
previous rugs were not delivered on time.
[^49]

(And firms were informed of this arrangement.) The experiment protocol simply
required Hamis to offer an
*initial*
order to the treatment firm.
[^50]

Hamis was not allowed to allocate any orders to control firms and we maintained
a project coordinator and survey team in Fowa to ensure that the protocols
were followed.
[^51]

Thus, the intervention provided treatment firms with the opportunity to
produce rugs for the export market.

Sample, Randomization and Take up
\label{subsec:Sample,-Randomization}


Since there is no official census of firms that manufacture rugs in Fowa,
we hired a Fowa-based NGO to go street by street looking for rug-making
firms in homes and workshops that satisfied the following characteristics:
a) had fewer than 5 employees; b) worked on their own account (meaning
that they bought their own inputs when an order required); c) had never
previously worked with Hamis, and d) were willing to participate in a research
experiment on exporting.
We stopped the recruitment drive when the NGO had identified 300 firms.
As there are far more than 300 rug-making firms in Fowa, the first recruitment
drive provided us with a sample of firms in central areas of Fowa that
were operating and observable at the time of recruitment.
This exercise produced a list of 303 firms, which we refer to as
"

Sample 1
"

.

Firms specialize in one of four rug types.
We stratified the sample on the type of rug produced and the loom size.
Within each stratum we randomized firms into treatment and control using
a random number generator and strata that contained an odd number of firms
were assigned one more firm to control than to treatment.
The first two rows of columns 2, 3, 4 and 6 in Table

show the total number of firms by rug type and treatment status for Sample
1.
The third row of Table

shows the takeup status for these firms.
As anticipated by our decision to stratify along this dimension, takeup
rates varied greatly depending on the firm's primary rug type.
For goublain and tups producers, the two rug types for which we obtained
no orders, takeup rates were 10 and 19 percent, respectively.
We expected low take up in these strata since these firms do not typically
produce duble or kasaees rugs.
Nevertheless, we attempted to treat these firms and found that very few
were willing to switch rug types.


In follow-up round 2, we asked all treated firms who did not take up why
they did not take up.
Table

presents these results.
The main reason for refusals among goublain and tups firms was that the
export rug order
"

was not the suitable rug type
"

.
While this could mean that either the rug category (i.e.
duble, tups, kasaees, goublan) was not suitable, or that the export rugs
had sufficiently different specifications such that firms perceived them
to be a different rug type, we take this as confirmatory evidence.

Table

shows that among kasaees and duble rug producers in Sample 1, take up was
26 and 38 percent, respectively.
These takeup rates were still relatively low.
As we detail in Section

, between December 2010 and May 2012 we were able to generate a small number
of export orders for kasaees and duble rugs (see Figure

).
However, given the small number of export orders, we were unable to approach
treatment firms in Sample 1 with the opportunity to produce the full 110m

in one go.
Instead, we had to sequentially offer smaller orders of 10-20m

, or about two weeks of work.
Because these orders were small, many firms were unwilling to work with
us: Table

shows that the most common reason why duble firms did not take up was an
unwillingness to jeopardize their existing relationships with other intermediar
ies for a small amount of work, and that the export order was not the suitable
rug type (presumably because of the very different export specifications).


Many kasaees producers were unwilling to accept the export order for the
same reason (i.e.
because there are several distinct types of kasaees and the particular
rug they were asked to produce was different from the kasaees rugs they
usually make).

From March 2013, Hamis' major buyers offered assurances that they would
continue to place duble rug orders for the foreseeable future so it became
feasible to offer the opportunity to produce 110m

in one go.
Since this was the intended treatment in our experimental protocol, we
conducted a second round of recruitment for firms that specialized in duble
rugs.
At this stage of the project, we had an experienced Fowa-based team who
were able to locate an additional 140 duble firms in Fowa that had been
not been found during the initial recruitment drive in July 2011.
To our knowledge, these were all the remaining duble rug producers in Fowa
with the exception of ten firms who refused to participate.
We refer to these additional firms as
"

Sample 2
"

and again stratified these firms by loom size.
We randomized 35 firms into the treatment group using a random number generator.
The choice of 35 treatment firms for Sample 2 was dictated both by Hamis'
constraints on the number of firms it could work with, and our desire to
ensure that the full 110m

could be offered to each treatment firm.
Column 5 of Table

reports treatment and takeup statistics for Sample 2.
Consistent with the low take up in Sample 1 being due to the inability
to offer the full 110m

packet in one go, 32 out of 35 treatment firms in Sample 2 took up the
opportunity to produce export orders.
[^52]

We combine the duble strata from Sample 1 with the firms in Sample 2—who
are exclusively duble firms—to form the
*Joint Sample*
of 219 firms (74 in treatment, 47 of which took up).
This is almost every duble producer in Fowa which mitigates concerns of
internal validity that would arise if we focused only on the second sample
that was treated with the full 110

initial order in one go.
This is essentially all the duble producers in Fowa willing to identify
themselves to our surveyors (minus ten firms who refused to participate).


By focusing on the Joint Sample instead of only Sample 2 (the sample that
received the treatment closest to the one we initially intended) we mitigate
internal validity concerns arising from unobservable differences between
the two samples.


(Although we secured some initial orders for kasaees rugs, we exclude them
from the analysis because were unable to secure further export orders for
this type of rug and so could not offer any follow-up orders for the 5
firms who took up.)

As noted above, the experiment protocol allowed Hamis Carpets to end their
relationship with firms that did not produce the 110m

packet at the expected quality level and in a timely manner.
The 5th row of Table

reports the number of
"

successful
"

takeup firms, defined as those who produced 110m


and then received subsequent orders from Hamis.
Only 3 treatment firms (all non-duble firms in Sample 1) failed to secure
additional orders from Hamis after the initial treatment.
Two of the firms were unable to manufacture the export orders successfully
while the other firm had a falling out with the owner of Hamis.

Samples 1 and 2 do differ on some observables.
On average, Sample 2 firms have higher profits and consumption, although
similar quality levels and output.
These differences may be due to the fact that recruitment was carried out
sequentially, so that Sample 2 firms were more difficult to find or less
centrally located.
However, there are no significant differences in observables between treatment
and control groups within either sample as can be seen in Table

, or in the Joint Sample of duble firms as can be seen in Table

.
We also find that the estimated treatment-on-the-treated (TOT) effects
do not differ across samples (we compare TOTs since ITTs will vary due
to differences in take up rates).
Out of the 44 comparisons of TOTs between the two samples, only three are
statistically significant at the 10 percent level and one is statistically
significant at the 5 percent level (in line with the number of type 1 errors
we would expect).

Protocol for Quality Lab
\label{ss_append-protocol-lab}

In June 2014, we rented a workshop with a loom and invited the primary weaver
in all the treatment and control firms in the Joint Sample to come to the
workshop to produce an identical rug.
The rug specifications mimicked a popular rug design sold at mid-tier domestic
retail outlets in Egypt (a 140cm by 70cm rug with a desired weight of 1750g,
and the master artisan assigned a difficulty rating of 3—see Figure

for the actual design).
Prior to the arrival of the firm's weaver (all of whom were the firm's
owners


), our field team prepared the loom for weaving with the help of the master
artisan.
We provided all the necessary thread, reed and heddle inputs.
Thus, each firm used an identical loom, an identical loom setup, and identical
inputs to manufacture an identical rug.

This setup was not unusual.
Some large intermediaries in Fowa have workshops where they bring producers
to and pay them to weave.
Nonetheless, our Quality Lab differed in that we invited firms to produce
a single rug, instead of many meters of rugs as is the market norm.
To mitigate potential Hawthorne effects, where the firms act differently
because they know they are being analyzed, we excluded Hamis Carpets from
being involved in administering the Quality lab, from participating in
contacting the firms, and from encouraging them to participate.
Instead, we hired a new staff member to administer the Quality Lab, and
he was instructed to invite sample firms to produce a rug at the rented
workshop for a LE70 payment, which is about three times the average price
for a rug of these specifications.
The firms were told that the reason why this lucrative opportunity existed
was because there was a buyer in Cairo that was interested in making a
one-time purchase of one rug from each of the many producers in Fowa.
This claim was truthful, as one of the authors was in Cairo at the time,
serving the role of that buyer.
There was no mention of Hamis Carpets and no discussion of whether performance
in the lab would affect future orders from Hamis.

The firms were shown to the Quality Lab and provided the design and specificatio
ns of the rug.
The firms were not incentivized to produce quality in any way, nor were
they encouraged to produce quickly or slowly, or given a time limit.
They also were not given any reason to believe that their performance on
this task would impact future opportunities to produce for this buyer,
or given any information about the desired quality of the rug.
They were not even told they were being evaluated in any way.
Firms were simply informed that as long as they completed the rug they
would be given LE70 as promised.
Our protocols directed the new staff member to interact with the treatment
and control group firms in an identical manner; from the way they were
recruited, to how they were told about the LE70 payment, and to the instruction
s they were given about what type of rug to produce.

The total time taken to weave the rug was recorded


and after the rug was completed, the staff member recorded the length,
height and weight of the rug, and tagged each rug with a new set of firm
identification numbers.
After all firms had manufactured the rugs, they were sent to the master
artisan to score the quality levels without revealing the master artisan
the identity of the firm that made them (only the anonymized identification
number).
In addition, the rugs were sent to Professor Fayrouz Al-Gamal, a Professor
of Handicraft Science from Domietta University, to provide an independent
set of quality scores.
Professor Al-Gamal has been the chair of the
"

Spinning, Weaving and Knitting Department
"

at the Domietta University since 2013 and is an expert in jacquard knitting
techniques.
He has contributed technical chapters to eight published books, and presented
his work at a dozen academic conferences in the last five years.

 \thefigure{E. figure}

 \thetable{E. table}

Profit Measures
\label{sec:Profits-Measures}

Following
de Mel et al. (2009)

, we construct four measures of profits.
The first measure directly records reported profits from the previous month.
Firms were asked:
"

What was the total income from the rug business last month after paying
all expenses (inputs, wages to weavers but excluding yourself).
That is, what were your profits from this business last month?
"

The second measure constructs profits from two survey questions that ask
firms to report their total revenues and total costs from the previous
month.

The third measure constructs profits from the production modules that contain
detailed information on prices and quantities of inputs and outputs, which
we refer to as
"

constructed profits
"

.
This measure is valuable if there is less noise in recalling prices and
quantities than total revenues and expenses.
This measure is also free of the concern that firms might use business
expenses for household consumption (or use business revenues to pay for
household expenses) that may be an issue for the other two measures.

The fourth measure is based on a hypothetical question that asks firms how
much they would earn from selling a specific quantity of inputs.
Specifically, we construct
"

hypothetical profit
"

by asking firms how much it would cost to purchase 25 kilograms of the
thread they used in the previous month, how long it would take to weave
this output, and how much they would earn from selling the output.
Although not the realized profits of the firm, this measure alleviates
potential concerns regarding the timing of when revenues are earned and
costs are incurred, and serves as a check against the other three profit
measures.

 \thefigure{F. figure}

 \thetable{F. table}


Additional Tables and Figures
\label{sec:Additional-Tables-and}


**Table: t_low**


    Goublain	Firms Tups	Firms Kasaees	Firms Duble	Firms All	Firms
    Reasons	for	Refusal N % N % N % N % N %
    (Agreed) 3 6.1 6 14.3 5 26.3 15 38.5 28 18.8
    Risk	relationship	with	current	intermediary 2 4.1 1 2.4 2 10.5 7 17.9 12 8.1
    Price	was	too	low 2 4.1 1 2.4 2 10.5 3 7.7 9 6.0
    Left	industry	or	passed	away 2 4.1 3 7.1 3 15.8 5 12.8 13 8.7
    Export	order	not	suitable	rug	type 39 79.6 30 71.4 6 31.6 7 17.9 82 55.0
    Refused	contact	with	survey	team 1 2.0 1 2.4 1 5.3 2 5.1 5 3.4
    Total 49 100 42 100 19 100 39 100 149 100
    Notes: Table reports the reasons for refusing treatment orders among Sample 1 firms as reported in the second survey round (April-May 2012). As of the
    second survey round, 28 firms had agreed to take orders. Since that time, an additional duble firm, two additional goublain firms and two additional tups
    firms	have	also	taken	orders	resulting	in	a	total	of	33	Sample	1	firms	takeing	up.


**Table: \label{t_hamismargins**


    Revenue	and	Expenses,	Egyptian	Pounds	per	m2
    Domestic	Orders Export	Orders
    Material	Expenses 30 40
    Payments	to	Producers 25 40
    Shipping	Costs 0 40
    Price	Received 60 160
    Markup 9% 33%
    Notes: Table reports Hamis Carpets' cost structure on foreign and domstic rugs. Numbers reported in Egyptian Pounds per square meter. The final row
    indicates	self-reported	markups	on	domestic	and	export	orders.


**Table: \label{t_attrition**


    Control Group
    Mean Treatment Obs.
    Attrition in All Follow Up Survey Rounds 0.11 0.00 815
    (0.01) (0.02)
    Attrition in Round 1, Sample 1 (Nov-Dec 2011) - -
    Attrition in Round 2, Sample 1 (April-May 2012) 0.15 -0.05 79
    (0.06) (0.08)
    Attrition in Round 3, Sample 1 (Sept-Dec 2012) 0.10 0.03 79
    (0.05) (0.07)
    Attrition in Round 4, Sample 1 (March-Apr 2013) 0.18 0.03 79
    (0.06) (0.09)
    Attrition in Round 1, Sample 2 (May-June 2013) 0.05 -0.03 140
    (0.02) (0.04)
    Attrition in Round 5, Sample 1 (July-Oct 2013) 0.20 0.05 79
    (0.07) (0.10)
    Attrition in Round 2, Sample 2 (Nov-Dec 2013) 0.02 0.01 140
    (0.01) (0.03)
    Attrition in Round 6, Sample 1 (Jan-March 2014) 0.28 0.03 79
    (0.07) (0.10)
    Attrition in Round 3, Sample 2 (May-June 2014) 0.07 -0.03 140
    (0.02) (0.04)
    Notes: Table reports attrition across all follow up survey rounds and then reports attrition seperately by round, in chronological order. Attrition is
    defined as a firm not having answered the survey in a particular round. Statistics regarding Sample 1 only report attrition for the Sample 1 duble
    firms that appear in the Joint Sample. As discussed elsewhere, we do not include Round 1 data for Sample 1 due to evidence of enumerator
    misconduct. Significance * .10; ** .05; *** .01.


**Table: \label{t_spillovers**


    Sum	of	Inverse	Distance Sum	of	Inverse	Distance
    to	Treatment	Firms Squared	to	Treatment Marginal	Effect R-Squared Obs.
    Firms
    Ever	Exported -0.2 16.3 6.8 0.12 128
    (0.9) (40.2) (16.4)
    Direct	Log	Monthly	Profits -0.9 0.3 -0.8 0.26 368
    (1.3) (43.2) (17.2)
    Direct	Log	Profits	per	Hour -0.5 -40.5 -17.7 0.18 368
    (1.4) (47.5) (19.0)
    Log	Output	per	Hour -1.3 37.3 19.1 0.08 427
    (1.0) (36.9) (16.4)
    Stacked	Quality 0.6 -7.1 -2.4 0.06 4,408
    (0.4) (15.9) (6.4)
    Log	Unadjusted	TFP -1.5 103.0 * 42.0 * 0.22 418
    (1.9) (59.3) (23.9)
    Log	Adjusted	TFP -1.7 * 55.4 21.9 0.10 421
    (1.0) (42.5) (17.3)
    Notes: Table reports results from regressing the outcome variable in each row on the sum of the inverse distance between control firms and all treatment
    firms and the sum if the inverse distance squared to all treatment firms (with distance measured in meters). The third column shows the marginal effect of
    distance on the outcome based on the results from the regression. Regressions include round and strata fixed effects and control for baseline values. The
    stacked	quality	regression	includes	quality	metric	fixed	effects.		Standard	errors	clustered	at	the	firm	level.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_fixedeffects**


    ITT TOT R-squared Obs.
    (1) (2) (3) (4)
    Ever	Exported 0.53 *** 0.74 *** 0.71 410
    (.12) (.10)
    Direct	Log	Monthly	Profits 0.30 ** 0.47 *** 0.57 874
    (.12) (.16)
    Direct	Log	Profits	per	Hour 0.28 ** 0.44 *** 0.48 874
    (.12) (.16)
    Log	Stacked	Quality 1.04 *** 1.58 *** 0.61 8,931
    (.10) (.09)
    Log	Output	per	Hour -0.18 -0.31 0.04 901
    (.12) (.20)
    Log	Unadjusted	TFP -0.26 ** -0.44 ** 0.04 890
    (.12) (.20)
    Log	Specification-Adjusted	Stacked	Quality 0.71 *** 0.88 *** 0.30 8,389
    0.09 0.08
    Log	Specification-Adjusted	TFP 0.26 ** 0.43 ** 0.15 298
    (.12) (.20)
    Notes: Table reports treatment effects on eight main outcomes for the Joint Sample of duble firms using firm fixed effects instead of controlling for the
    baseline	value	of	the	variable.	All	regressions	include	firm	and	round	fixed	effects,	and	cluster	standard	errors	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_dynamics**


    Log Direct Profits Log Output per Log Unadjusted Log Specification- Log Specification-
    Log Direct Profits per Hour Hour TFP Adjusted TFP Adjusted Quality
    (1) (2) (3) (4) (5) (6)
    Treatment 0.28 *** 0.23 *** -0.26 ** -0.29 *** 0.19 * 0.44 ***
    (0.08) (0.07) (0.11) (0.10) (0.10) (0.06)
    x 2nd Survey Round -0.03 -0.06 0.16 0.18 0.13 0.01
    (0.10) (0.09) (0.12) (0.12) (0.11) (0.06)
    x 3rd Survey Round -0.07 -0.12 -0.05 -0.08 -0.01 0.01
    (0.10) (0.09) (0.13) (0.12) (0.11) (0.07)
    x 4th Survey Round -0.07 -0.04 -0.26 -0.31 -0.33 * -0.10
    (0.16) (0.15) (0.24) (0.23) (0.18) (0.11)
    x 5th Survey Round 0.27 * 0.27 * 0.17 0.13 0.07 -0.21 *
    (0.16) (0.15) (0.24) (0.21) (0.15) (0.11)
    R-squared 0.22 0.15 0.18 0.27 0.13 0.18
    Observations 573 573 687 674 669 6,860
    Notes:  Table reports treatment effects interacted with dummies for each round of data collection. The regressions control for baseline values of the
    dependent variable, and include round and strata fixed effects. Standard errors are clustered by firm. Significance * .10; ** .05; *** .01.


**Table: \label{t_visit_summary**


    (1)
    Number of Visits 11.0
    (2.57)
    Length of Visit (in minutes) 27.6
    (4.88)
    Discussed technique (as opposed to pointing out flaws) 90.3%
    Observations 44
    Discussed Metric Discusssed Technique
    (1A) (2A)
    Corners 31.8% 100.0%
    Waviness 20.5% 100.0%
    Weight 54.5% 92.9%
    Touch 11.4% 100.0%
    Packedness 20.5% 93.8%
    Warp Thread Tightness 47.7% 78.9%
    Firmness 31.8% 100.0%
    Design Accuracy 50.0% 96.2%
    Warp Thread Packedness 22.7% 75.0%
    Observations 44
    Notes: Table summarizes the data collected about knowledge flows between the intermediary and the firms. All firms were visited at least 7
    times, and the top panel reports the average length of each visit in minutes (with standard deviations in parantheses), and the proportion of
    interactions that discuss technique, rather than simply pointing out flaws. The bottom panel reports the proportion of firms that report
    discussing the quality metric with the intermediary and, conditional on discussing that metric, the proportion of firms that report that the
    intermediary discussed techniques to improve along that dimensions rather than simply pointing out flaws. Note that these data were collected
    before two take-up firms in Sample 2 began producing for export.


**Table: \label{t_hedonics**


    Log	Profits	Per	Hour
    (1)
    Corners 0.08
    (0.06)
    Waviness -0.044
    (0.06)
    Weight -0.027
    (0.05)
    Touch 0.189 **
    (0.07)
    Packedness -0.150 **
    (0.07)
    Warp	Thread	Tightness 0.173 **
    (0.08)
    Firmness -0.103
    (0.08)
    Design	Accuracy 0.106 *
    (0.05)
    Warp	Thread	Packedness 0.032
    (0.06)
    P-Value	of	Joint	F-Test 0.013
    R-squared 0.587
    Observations 563
    Notes: Table reports the regression of log profits per hour on the nine quality metrics that are recorded in Step 2. The regression is run on
    all firms producing for the domestic market (the non-takeup firms). Regressions also include strata and round fixed effects. Standard errors
    are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.
    Notes: Table reports the regression of log profits per hour on the nine quality metrics that are recorded in Step 2. The regression is run on
    all firms producing for the domestic market (the non-takeup firms). Regressions also include strata and round fixed effects. Standard errors
    are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


\begin{sidewaysfigure[H]

Semi-Parametric Learning Curves for Takeup Firms}


sidewaysfigure[t]

Learning Curves using High-Frequency Order-Book Data for Takeup Firms}


sidewaysfigure[H]

Semi-Parametric Learning Curves using High-Frequency Order-Book Data for
Takeup Firms}


**Figure: f_learning_cdf**


**Figure: \label{f_learning_days**


**Figure: \label{f_fray**


\renewcommand{ \thefigure{F. figure}

 \thetable{F. table}

Results for Sample 1 duble firms, Sample 2 firms, and Sample 1 non-duble
firms
\label{s_sample12}

Tables

-

and Figures

-

in this appendix repeat the key tables and figures from the main text but
now splitting the Joint Sample into Sample 1 duble firms and Sample 2 firms
(recall all Sample 2 firms are duble firms).
Table

reports the treatment effects for the non-duble firms in Sample 1 who we
were not able to provide with sufficient orders in their specialist rug
type.
[^53]


**Table: t_balance-12**


    Sample	1 Sample	2
    Control Control
    Group	Mean Treatment Obs. Group	Mean Treatment Obs.
    Panel	A:	Household	Characteristics
    Age 51.4 -1.5 79 50.7 2.8 139
    (1.1) (2.2) (1.0) (2.2)
    Number	of	years	in	rug	business 39.3 -2.0 77 36.8 1.9 136
    (1.1) (2.3) (1.1) (2.5)
    Illiterate? 0.70 0.14 79 0.59 0.07 135
    (0.05) (0.10) (0.04) (0.10)
    Household	size 4 .5 -0.1 79 4 .1 0.1 140
    (0.2) (0.3) (0.1) (0.3)
    Household	income 1,100.0 -121.0 79 1,090.0 226.0 140
    (137.0) (278.0) (120.0) (340.0)
    Digit	Span	Recall 5 .0 -0.1 72 6 .3 0.5 132
    (0.1) (0.2) (0.1) (0.3)
    Panel	B:	Firm	Characteristics
    Price	per	square	meter 38.7 13.0 79 25.4 2.1 139
    (8.6) (17.5) (1.8) (3.9)
    Direct	monthly	profits	from	rug	business 2 66 40.2 79 8 62 -16.6 139
    (75.3) (153.0) (39.5) (86.0)
    Reported	monthly	profits	from	rug	business 7 14 -8.8 78 8 58 -11.5 139
    (79.4) (162.0) (39.9) (86.2)
    Hours	worked	last	month 2 08 -5.7 79 2 69 1.3 139
    (11.5) (23.3) (5.1) (10.8)
    Number	of	employees 1.25 0.00 79 1.00 139
    (0.08) (0.17) - -
    Total	produced	last	month	(m2) 61.3 7.11 79 43.6 0.3 139
    (10.9) (22.15) (2.4) (5.8)
    Ever	exported? 0.03 0.00 79 0.17 0.03 140
    (0.02) (0.04) (0.03) (0.08)
    Average	Quality 0.22 -0.18 ** 78 0.19 -0.09 140
    (0.05) (0.08) (0.03) (0.06)
    Constructed	monthly	profits	from	rug	business 8 65 -35.6 139
    (42.6) (87.4)
    Hypothetical	monthly	profits	from	rug	business 3 65 -3.6 139
    (26.5) (57.7)
    Joint	F-test 1.01 1.37
    Notes: Table presents baseline balance for Sample 1 duble-firms (left panel) and the Sample 2 (right panel). Each row is a regression of the variable on a constant,
    treatment dummy and strata fixed effects; the constant (control group mean) and treatment dummy are reported. The last row reports the F-test for a regression
    of the treatment dummy on all 14 baseline balance variables. Profits and price are winsorized at the 1st and 99th percentile to trim outliers (without winsorizing,
    the sample still remains statistically balanced between treatment and control groups). The standard error and treatment dummy for "number of employees" in
    Sample 2 are missing because every firm had one employee, inclusive of the owner. Constructed profits and hypothetical profits were not collected at baseline for
    sample	1.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_firststage-12**


    Sample	1 Sample	2 P-Value	of
    ITT TOT ITT TOT TOT
    (1) (2) (3) (4) Comparison
    Indicator	for	Ever	Exported 0.31 ** 0.79 *** 0.68 *** 0.75 *** 0.90
    (0.12) (0.25) (0.07) (0.07)
    R-squared 0.12 0.34 0.45 0.49
    Control	Mean 0.19 0.19 0.20 0.20
    Observations 59 59 132 132
    Notes: Table regresses an indicator for if a firm has ever knowingly produced rugs for export markets on indicators for treatment (columns 1 and 3) or takeup
    (columns 2 and 4). The question was asked in Round 5 for Sample 1 and Round 3 for Sample 2. The TOT regressions instrument takeup with treatment. The last
    column reports the p-value from the statistical test of equivalence of the TOT coefficients between the two samples. The regressions control for baseline values
    of	the	dependent	variable,	and	include	round	and	strata	fixed	effects.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_profits-12**


    Panel	A:	Profits	(in	Month	Prior	to	Survey)
    Sample	1 Sample	2
    Control Control P-Value	of	TOT
    ITT TOT Mean Obs. ITT TOT Mean Obs. Comparison
    Outcome	Variable: (1) (2) (3) (4) (5) (6) (7) (8) (9)
    Log	Direct	Profits 0.27 ** 0.83 ** 890 198 0.25 *** 0.30 *** 951 375 0.11
    (.11) (.33) (.06) (.07)
    Log	(Reported	Revenues	-	Reported	Costs) 0.22 ** 0.71 ** 892 269 0.23 *** 0.28 *** 955 375 0.21
    (.11) (.33) (.05) (.06)
    Log	(Constructed	Revenues	-	Constructed	Costs) 0.14 0.48 940 310 0.24 *** 0.29 *** 957 375 0.59
    (.10) (.34) (.05) (.07)
    Log	Hypothetical	Profits 0.39 ** 1.29 ** 465 314 0.36 *** 0.44 *** 586 373 0.19
    (.18) (.64) (.10) (.12)
    Panel	B:	Profits	per	Owner	Hour	(in	Month	Prior	to	Survey)
    Sample	1 Sample	2
    Control Control P-Value	of	TOT
    ITT TOT Mean Obs. ITT TOT Mean Obs. Comparison
    Outcome	Variable: (1) (2) (3) (4) (5) (6) (7) (8) (9)
    Log	Direct	Profits 0.24 ** 0.74 ** 3.48 198 0.17 *** 0.21 *** 3.56 375 0.09
    (.10) (.30) (.05) (.06)
    Log	(Reported	Revenues	-	Reported	Costs) 0.21 ** 0.66 ** 3.49 262 0.15 *** 0.19 *** 3.58 375 0.12
    (.10) (.30) (.05) (.06)
    Log	(Constructed	Revenues	-	Constructed	Costs) 0.15 * 0.52 * 3.50 309 0.16 *** 0.19 *** 3.58 375 0.28
    (.09) (.30) (.05) (.06)
    Log	Hypothetical	Profits 0.29 ** 0.98 ** 4.80 314 0.21 *** 0.26 *** 6.01 373 0.07
    (.11) (.38) (.07) (.08)
    Notes: Table reports treatment effects on different measures of real profits in the month prior to the date of the survey, all measured in logs, seperately for Sample 1
    duble-firms and Sample 2. See text for descriptions of each measure. Dependent variable in Panel A is profits. Dependent variable in Panel B is profits per owner hour.
    Owner hours include the hours of family member production when recorded. The regressions control for baseline values of the dependent variable, and include round
    and strata fixed effects. Control group means are reported in levels in Egyptian pounds (LE) in Panel A and LE/hour in Panel B. The TOT regressions instrument takeup
    with	treatment.	Standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_components-12**


    Panel	A:	Components	of	Profits
    Sample	1 Sample	2
    Control Control P-Value	of	TOT
    ITT TOT Mean Obs. ITT TOT Mean Obs. Comparison
    Outcome	Variable: (1) (2) (3) (4) (5) (6) (7) (8) (9)
    Log	Output	Prices	(LE/m2) 0.40 ** 1.36 ** 26.4 315 0.46 *** 0.56 *** 29.3 376 0.23
    (0.18) (0.65) (0.10) (0.12)
    Log	Output	(m2) -0.30 * -1.01 * 74.7 301 -0.22 ** -0.27 *** 57.9 375 0.21
    (0.16) (0.59) (0.09) (0.10)
    Log	Monthly	Hours	Worked 0.01 0.03 265.0 303 0.08 *** 0.10 *** 272.0 375 0.57
    (0.03) (0.10) (0.02) (0.03)
    Number	of	Employees 0.01 0.01 1.0 319 0.00 0.00 1.0 376 .
    (0.01) (0.01) . .
    Log	Number	of	Looms 0.01 0.04 1.2 318 -0.06 *** -0.07 *** 1.1 376 0.62
    (0.07) (0.22) (0.01) (0.02)
    Log	Warp	Thread	Ball	(kg) -0.01 -0.03 5.2 311 0.13 ** 0.15 ** 6.3 377 0.13
    (0.20) (0.68) (0.05) (0.06)
    Panel	B:	Inputs
    Sample	1 Sample	2
    Control Control P-Value	of	TOT
    ITT TOT Mean Obs. ITT TOT Mean Obs. Comparison
    Outcome	Variable: (1) (2) (3) (4) (5) (6) (7) (8) (9)
    Log	Weft	Thread	Price	(LE/m2) 0.16 0.62 12.3 188 0.23 *** 0.29 *** 13.1 376 0.54
    (0.15) (0.53) (0.04) (0.05)
    Log	Warp	Thread	Price	(LE/m2) -0.03 -0.10 19.8 309 -0.03 -0.04 17.0 376 0.74
    (0.05) (0.17) (0.03) (0.04)
    Log	Weft	Thread	Quantity -0.25 -0.85 130.0 302 -0.13 -0.16 97.8 375 0.26
    (0.17) (0.60) (0.09) (0.10)
    Log	Warp	Thread	Quantity -0.01 -0.03 17.7 311 0.08 0.10 17.8 375 0.85
    (0.20) (0.68) (0.09) (0.11)
    Notes: Table reports results seperately for Sample 1 duble-firms and Sample 2. Panel A reports treatment effects on output prices and quantities, hours, number of
    employees (inclusive of owner), looms and the size of the warp thread ball (which is a proxy for the length of the production run), all measured in logs except number
    of employees. Panel B analyzes input prices and quantities, all measured in logs. The TOT regressions instrument takeup with treatment. Hours worked are calculated
    using average daily hours and number of days worked last month. Control group means are reported in levels. The regressions control for baseline values of the
    dependent variable, and include round and strata fixed effects. P-values from testing equivalence of the TOTs from each sample are reported in the last column. We
    cannot estimate the ITTs or TOTs for "number of employees" as there is no variation in sample 2. Standard errors are clustered by firm. Significance * .10; ** .05; ***
    .01.


**Table: \label{t_quality-12**


    Panel	A:	Quality	Metrics
    Sample	1 Sample	2
    Control ITT TOT Control ITT TOT
    Mean (1) (2) Mean (3) (4)
    Corners 2.89 0.66 *** 1.73 *** 3.00 1.38 *** 1.69 ***
    (0.21) (0.41) (0.13) (0.08)
    Waviness 2.89 0.67 *** 1.76 *** 3.02 1.36 *** 1.66 ***
    (0.20) (0.36) (0.13) (0.08)
    Weight 2.90 0.67 *** 1.75 *** 3.14 1.32 *** 1.60 ***
    (0.20) (0.38) (0.12) (0.09)
    Touch 3.14 0.24 *** 0.67 *** 3.11 0.54 *** 0.65 ***
    (0.08) (0.18) (0.08) (0.06)
    Packedness 3.17 0.39 ** 1.32 *** 3.08 1.38 *** 1.68 ***
    (0.15) (0.36) (0.12) (0.08)
    Warp	Thread	Tightness 3.04 0.43 *** 1.38 *** 3.05 1.24 *** 1.51 ***
    (0.15) (0.35) (0.12) (0.09)
    Firmness 2.99 0.34 ** 1.19 *** 2.97 1.43 *** 1.75 ***
    (0.14) (0.35) (0.13) (0.08)
    Design	Accuracy 3.28 0.36 ** 1.21 *** 3.12 1.22 *** 1.48 ***
    (0.15) (0.35) (0.12) (0.10)
    Warp	Thread	Packedness 2.90 0.66 *** 1.71 *** 3.10 1.33 *** 1.64 ***
    (0.20) (0.40) (0.13) (0.09)
    Inputs 3.09 0.44 *** 1.51 *** 3.06 1.37 *** 1.66 ***
    (0.15) (0.38) (0.11) (0.09)
    Loom 2.03 0.02 0.06 2.02 0.04 0.05
    (0.02) (0.06) (0.04) (0.04)
    R-squared 0.32 0.50 0.57 0.66
    Observations 2,765 2,765 4,120 4,120
    Panel	B:	Stacked	Quality	Metrics
    Sample	1 Sample	2
    Control ITT TOT Control ITT TOT
    Mean (1) (2) Mean (3) (4)
    Stacked	Quality	Metrics 2.94 0.41 *** 1.26 *** 2.97 1.14 *** 1.39 ***
    (0.13) (0.27) (0.10) (0.06)
    P-Value	of	TOT	Comparison 0.63
    R-squared 0.29 0.44 0.52 0.60
    Observations 2,765 2,765 4,120 4,120
    Notes: The table reports the estimated impacts seperately for Sample 1 duble-firms and Sample 2. Panel A stacks the quality metrics and interacts treatment (ITT)
    or takeup (TOT) with a quality-metric indicator variable. The coefficients on the interactions provide the treatment effects seperately for each metric. The TOT
    instruments takeup (interacted with quality metric) with treatment (also interacted with quality metric). Each regression includes baseline values of the quality
    metric, strata and round fixed effects, and each of these controls is interacted with a quality-metric indicator variable. Panel B constrains the treatment effects to
    be equal across quality metrics; these regressions include baseline values, strata and round fixed effects. Control group means are reported in levels. Standard
    errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_outputhr-12**


    Sample	1 Sample	2
    Control Control P-Value	of	TOT
    ITT TOT Mean Obs. ITT TOT Mean Obs. Comparison
    (1) (2) (3) (4) (5) (6) (7) (8) (9)
    Log	Output	Per	Hour -0.24 -0.79 0.30 311 -0.24 *** -0.29 *** 0.24 376 0.47
    (0.15) (0.53) (0.09) (0.10)
    Log	Unadjusted	TFP -0.30 * -1.00 * 0.58 299 -0.26 *** -0.32 *** 0.43 375 0.32
    (0.15) (0.55) (0.09) (0.10)
    Notes: TFP Table reports treatment effects for the two productivity measures: log unadjusted output per labor hour (in m2/hour) and log unadjusted TFP, seperately
    for Sample 1 duble-firms and Sample 2. See text and Appendix A for the methodology used to obtain unadjusted TFP. The TOT specifications instrument takeup with
    treatment. Control group means are reported in levels. Regressions control for baseline values of the variable, round and strata fixed effects. Standard errors are
    clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


Quality and Productivity, Sample 1 and Sample 2 (Appendix to Table

)


**Table: t_lab-12**


    Panel	A:	Quality	Metrics
    Sample	1 Sample	2
    Master	Artisan Professor Master	Artisan Professor
    Control ITT TOT Control ITT TOT Control ITT TOT Control ITT TOT
    Mean (1A) (1B) Mean (2A) (2B) Mean (3A) (3B) Mean (4A) (4B)
    Corners 3.27 0.47 * 1.33 ** 3.41 0.14 0.39 3.22 0.89 *** 0.98 *** 3.27 0.40 ** 0.44 **
    (0.25) (0.56) (0.20) (0.52) (0.16) (0.15) (0.17) (0.18)
    Waviness 3.24 0.30 0.84 3.27 0.19 0.55 3.14 0.72 *** 0.79 *** 3.33 0.29 * 0.32 **
    (0.25) (0.56) (0.19) (0.45) (0.17) (0.17) (0.15) (0.16)
    Weight 3.62 0.27 0.78 3.89 0.53 ** 1.50 ** 3.59 0.85 *** 0.94 *** 3.53 0.62 ** 0.69 ***
    (0.20) (0.49) (0.22) (0.73) (0.16) (0.16) (0.24) (0.26)
    Packedness 3.32 0.31 0.88 * 3.31 0.07 0.21 3.28 1.09 *** 1.20 *** 3.26 0.43 *** 0.47 ***
    (0.22) (0.49) (0.19) (0.48) (0.15) (0.15) (0.14) (0.15)
    Touch 3.41 0.22 0.62 3.27 0.25 0.70 3.24 0.73 *** 0.80 *** 3.27 0.43 *** 0.47 ***
    (0.20) (0.45) (0.20) (0.46) (0.13) (0.13) (0.15) (0.16)
    Warp	Thread	Tightness 3.03 0.29 * 0.81 ** 3.36 -0.02 -0.06 2.99 0.66 *** 0.73 *** 3.28 0.43 *** 0.48 ***
    (0.16) (0.35) (0.19) (0.52) (0.10) (0.10) (0.15) (0.16)
    Firmness 3.32 0.22 0.62 3.24 0.074 0.21 3.16 1.04 *** 1.15 *** 3.22 0.44 *** 0.48 ***
    (0.26) (0.64) (0.19) (0.48) (0.14) (0.14) (0.15) (0.17)
    Design	Accuracy 3.75 0.30 0.85 * 3.55 -0.01 -0.02 3.61 0.68 *** 0.75 *** 3.41 0.45 *** 0.50 ***
    (0.18) (0.46) (0.21) (0.60) (0.14) (0.14) (0.12) (0.13)
    Warp	Thread	Packedness 3.16 0.51 ** 1.45 ** 3.28 0.14 0.39 3.00 1.12 *** 1.23 *** 3.17 0.57 *** 0.62 ***
    (0.25) (0.57) (0.19) (0.50) (0.16) (0.16) (0.15) (0.16)
    R-squared 0.11 0.31 0.16 0.18 0.31 0.36 0.10 0.10
    Observations 593 593 589 589 1,087 1,087 1,078 1,078
    Panel	B:	Stacked	Quality	Metrics
    Sample	1 Sample	2
    Master	Artisan Professor Master	Artisan Professor
    Control ITT TOT Control ITT TOT Control ITT TOT Control ITT TOT
    Mean (1A) (1B) Mean (2A) (2B) Mean (3A) (3B) Mean (4A) (4B)
    Stacked	Quality	Metric 3.35 0.32 * 0.91 ** 3.40 0.15 0.43 3.25 0.87 *** 0.95 *** 3.30 0.45 *** 0.50 ***
    (0.18) (0.40) (0.16) (0.40) (0.12) (0.11) (0.12) (0.13)
    P-Value	of	TOT	Comparison 0.92 0.88
    R-squared 0.09 0.30 0.14 0.22 0.29 0.34 0.09 0.09
    Observations 593 593 589 589 1,087 1,087 1,078 1,078
    Panel	C:	Objective	Metrics
    Sample	1 Sample	2
    Control ITT TOT Control ITT TOT P-Value
    Mean (1A) (1B) Mean (3A) (3B) of	TOT
    Length	Accuracy -4.41 0.70 1.98 -4.56 1.93 *** 2.12 *** 0.95
    (0.85) (2.17) (0.63) (0.70)
    Width	Accuracy -2.22 -0.21 -0.60 -2.32 0.43 0.48 0.48
    (0.52) (1.50) (0.34) (0.36)
    Weight	Accuracy -197.0 83.0 *** 236.0 ** -231.0 93.3 *** 103.0 *** 0.18
    (26.8) (93.4) (29.1) (30.6)
    Time	(in	minutes) 255.0 -22.10 * -62.70 * 243.0 5.52 6.1 0.05
    (12.1) (34.2) (7.4) (8.0)
    R-squared 0.87 0.81 0.83 0.83
    Observations 264 264 484 484
    Notes: Table reports ITT and TOT specifications using the 9 quality metrics from the quality lab, seperately for Sample 1 duble-firms and Sample 2. Panel A stacks the quality
    metrics and interacts treatment (ITT) or takeup (TOT) with a quality-metric indicator variable. The coefficients on the interactions provide the treatment effects seperately for
    each metric. The TOT instruments takeup (interacted with quality metric) with treatment (also interacted with quality metric). Panel B constrains the treatment effects to be
    equal across quality metrics. Columns 1A and 1B report scores from the master artisan. Columns 2A and 2B report scores from the Professor of Handicraft Science at Domietta
    University. Panel C reports objective accuracy measures which are calculated as the negative of the absolute error for that specification, so that a higher value indicates that the
    manufactured rug was closer to intended length (140 cm), width (70 cm) and weight (1750 g). It also includes the time spent to produce the rug in minutes. As in Panel A, these
    are run in a single regression by intereacting the objective measure with treatment or takeup. All regressions include interactions of strata fixed effects with quality-metric
    indicators,	and	standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


**Table: \label{t_dynamics-1**


    Log	Direct	Profits Log	Direct	Profits	per	Hour Log	Output	per	Hour
    Sample	1 Sample	2 Sample	1 Sample	2 Sample	1 Sample	2
    (1A) (1B) (2A) (2B) (3A) (3B)
    Treatment 0.36 0.24 *** 0.31 0.20 *** -0.24 -0.28 **
    (0.21) (0.08) (0.19) (0.07) (0.22) (0.12)
    x	2nd	Round	Since	Treatment -0.23 0.05 -0.17 -0.01 0.23 0.11
    (0.28) (0.08) (0.25) (0.08) (0.21) (0.14)
    x	3rd	Round	Since	Treatment -0.17 -0.03 -0.21 -0.09 -0.14 0.01
    (0.24) (0.09) (0.23) (0.09) (0.27) (0.13)
    x	4th	Round	Since	Treatment -0.16 -0.12 -0.28
    (0.25) (0.23) (0.26)
    x	5th	Round	Since	Treatment 0.16 0.16 0.14
    (0.23) (0.21) (0.29)
    R-squared 0.15 0.29 0.15 0.18 0.18 0.19
    Observations 198 375 198 375 311 376
    Log	Unadjusted	TFP Log	Specification-Adjusted	TFP Log	Specification-Adjusted	Quality
    Sample	1 Sample	2 Sample	1 Sample	2 Sample	1 Sample	2
    (1A) (1B) (2A) (2B) (3A) (3B)
    Treatment -0.31 -0.28 ** 0.19 * 0.37 *** -0.03 0.63 ***
    (0.20) (0.11) (0.10) (0.09) (0.05) (0.08)
    x	2nd	Round	Since	Treatment 0.33 0.08 0.13 -0.02 0.16 -0.03
    (0.24) (0.13) (0.11) (0.11) (0.10) (0.08)
    x	3rd	Round	Since	Treatment -0.17 -0.02 -0.01 0.03 0.34 *** -0.06
    (0.24) (0.12) (0.11) (0.12) (0.11) (0.08)
    x	4th	Round	Since	Treatment -0.29 -0.33 * 0.38 ***
    (0.25) (0.18) (0.10)
    x	5th	Round	Since	Treatment 0.15 0.07 0.27 ***
    (0.24) (0.15) (0.09)
    R-squared 0.31 0.18 0.13 0.16 0.13 0.26
    Observations 299 375 669 371 2,784 4,076
    Notes:		Table	reports	treatment	effects	interacted	with	dummies	for	each	round	of	data	collection,	separately	for	Sample	1	duble-firms	and	Sample	2.	The
    regressions	control	for	baseline	values	of	the	dependent	variable,	and	include	round	and	strata	fixed	effects.	Standard	errors	are	clustered	by	firm.		Standard
    errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


Information Flows and Quality Levels: Sample 1 and Sample 2 (Appendix to
Table

)


**Table: t_nondubs**


    TOT	for	Joint P-Value	for	TOT
    ITT TOT R-squared Obs. Sample Comparison
    (1) (2) (3) (4) (5) (6)
    Ever	Exported 0.03 0.22 0.28 90 0.76 *** 0.40
    (.09) (.63) (.07)
    Direct	Log	Monthly	Profits -0.01 -0.07 0.05 324 0.42 *** 0.52
    (.08) (.76) (.08)
    Direct	Log	Profits	per	Hour 0.00 -0.02 0.04 323 0.32 *** 0.64
    (.07) (.72) (.08)
    Log	Stacked	Quality 0.03 0.30 0.58 2830 1.35 *** 0.02
    0.04 (.45) (.08)
    Log	Output	per	Hour 0.05 0.65 0.69 373 -0.42 *** 0.49
    (.10) (1.30) (.16)
    Log	Unadjusted	TFP 0.03 0.38 0.65 351 -0.50 *** 0.62
    (.13) (1.76) (.16)
    Log	Specification-Adjusted	Stacked 0.03 0.36 0.23 2800 0.72 *** 0.35
    Quality 0.04 (.37) (.04)
    -0.04 -0.59 0.31 345 0.36 *** 0.77
    Log	Specification-Adjusted	TFP
    (.09) (1.27) (.12)
    Notes: Table reports treatment effects on eight main outcomes for the firms in Sample 1 who were not in the duble strata. The sample sizes vary across outcomes
    and in general are lower than for duble strata because after round 3, we stopped regular data collection on these strata because we did not secure sufficient
    overseas orders for these rugs and had budgetary constraints. in round 5 (essentially the endline), we collected data for the kasaees and tups strata, but did not
    collect data for the goublain strata which had the lowest takeup of all strata. Regressions include baseline values of the dependent variable and round and strata
    fixed	effects.	Standard	errors	are	clustered	by	firm.	Significance	*	.10;	**	.05;	***	.01.


\begin{sidewaysfigure[t]

Learning Curves, Sample 1 Takeup Firms (Appendix to Figure

)}


sidewaysfigure[t]

Learning Curves, Sample 2 Takeup Firms (Appendix to Figure

)}




---

## Notes

[^1]: One solution is to restrict attention to homogenous goods, such as concrete,
 block ice or gasoline (e.g.

FosterHaltiwangerSyverson2008

).
 This is unappealing for the study of learning-by-exporting since there
 is likely to be less scope for learning and there are fewer trading frictions
 in homogenous goods industries.

[^2]: For example, 
Clerides et al. (1998)

 and 
Bernard and Jensen (1999)

 conclude that firms self-select into export markets.
 In contrast, several papers using alternative approaches to deal with selection
 (e.g.
 matching estimators or instrumental variables) find some support for learning;
 see 
Loecker (2007)

, 
Park et al. (2010)

, 
Marin and Voigtlander (2013)

 and 
Loecker (2013)

.

(KellerJEL2004)

, 
(WagnerWorldEconomy2007)

 and 
Harrison and Rodriguez-Clare (2010)

 survey the literature.

[^3]: For example, see 
(Verhoogen (2008))

, 
Manova and Zhang (2012)

, 
(Crozet et al. (2012))

, 
(Brambilla et al. (2012))

, 
Hallak (2013)

, and 
Bastos et al. (2014)

.
 In contrast, 
Marin and Voigtlander (2013)

 find that rather than quality rising, marginal costs decline because Columbian
 firms invest in reducing production costs as they enter export markets.

[^4]: Papers that infer quality in these ways include 
Verhoogen (2008)

, 
Manova and Zhang (2012)

, 
(Khandelwal (2010))

, 
(Hallak (2011))

, and 
Feenstra and Romalis (2014)

.

[^5]: fn:Chinese

These numbers come from the data described by 
(Ahn et al. (2011))

 who show that smaller firms use intermediaries in order to avoid large
 fixed costs associated with directly exporting.

[^6]: Duble and tups rugs are the most common; kasaees rugs are the cheapest and
 woven from rags; goublan rugs are the most expensive and are works of art
 hung on walls.
 See Figure 
[f_rugpics]

 for examples.

[^7]: As discussed below, there are two baseline surveys that were run in July
 2011 and February 2013.
 The exchange rate on July 1, 2011 was LE5.94 and LE6.68 on February 1, 2013;
 we apply the average value, 6.31, throughout.

[^8]: The revenues received by the intermediary from these orders were significantly
 larger.
 As shown in Table 
[t_hamismargins]

 and explained further below, Hamis reports applying a 33 percent markup
 on export orders after paying for materials.

[^9]: It is standard practice in Fowa for intermediaries to provide design-specific
 reeds and heddles, and the cost to the intermediary does not depend on
 the design.
fn:heddle

[^10]: A project coordinator and Fowa-based survey team ensured that the protocols
 were followed.
 However, one control firm was incorrectly treated due to an error by Hamis.
 In the empirical analysis we make the most conservative assumption and
 keep this firm in the control group.

[^11]: fn:treat-heterog

Although the initial order is random, subsequent orders are not.
 Hence, treatment effects are potentially heterogenous if some firms are
 better able to take advantage of the initial opportunity than others.
 If there was a constraint on the total export orders, better firms may
 receive subsequent orders at the expense of worse firms.
 Under the reasonable assumption that learning is concave in export orders
 (an assumption supported by the learning curves we estimate below), this
 will lead us to underestimate the impacts of the experiment compared to
 a setting with unconstrained orders.
 To mitigate this concern, we restricted the size of the treatment group
 in Sample 2 to ensure that constraints on offering orders to all treatment
 firms would not be binding.

[^12]: It is important to note that although the two samples look very similar
 in terms of characteristics (see Table 
[t_balance-12]

), Sample 1 received a fundamentally different treatment—a treatment of
 multiple smaller export orders spread over time—from Sample 2.
 Nevertheless, other than the lower take up, results are similar across
 the two samples.
 Out of 43 comparisons of treatment-on-the-treated coefficients between
 the two samples, three are significantly different at the 10 percent level
 and one is significant at the 5 percent level (in line with the number
 of Type 1 errors we would expect).

[^13]: The Sample 1 baseline survey recorded 6 quality metrics to which we subsequently
 added 5 more metrics.
 Corners captures the straightness of where the rug edges meet.


 Waviness captures how flat the rug lies when placed on a hard surface.
 Weight captures how close the actual weight of the rug is to the intended
 weight.
 Touch reflects the feel of the rug.
 Packedness measures how well the rug holds together (poorly packed rugs
 have small holes).
 Warp thread tightness measures the tightness of the warp thread which helps
 determine how tightly held the weft thread is.
 Firmness measures the firmness of the rug when held.
 Design accuracy captures how accurate the design is to the intended pattern.
 Warp thread packedness measures how visible the warp thread is (it should
 not be visible at all).
 Inputs measures the quality of the input threads.
 Loom measures the quality of the loom.

[^14]: fn:spillovers

We note that any effects we find will be attenuated if there were spillovers
 to control firms.
 However, we find no support for geographic spillovers between treatment
 and control firms (see Table 
[t_spillovers]

).

[^15]: fn:ancova

Alternatively we could use all survey rounds, include firm fixed effects,
 and interact 
$Treatment_{i}$

 with a post-baseline dummy.
 We prefer our specification since if the dependent variable is measured
 with noise and not strongly autocorrelated, as is the case for business
 profits (an autocorrelation of 0.33 among our control firms), the fixed
 effects estimator will perform more poorly in an experimental study


 than the ANCOVA estimator in 
e_itt

; see 
McKenzie (2012)

.
 For comparison we also report the key results using firm fixed effects
 in Table 
[t_fixedeffects]

 and find that results are very similar.


 When the baseline value is missing for a firm, our specification drops
 that firm, leading the number of observations to vary across estimates.
 In cases where the baseline value is missing for an entire strata, we include
 the firm but code the missing value as zero.
 We find no evidence that missing values at baseline are correlated with
 treatment when testing the eight main outcomes of the paper (i.e.
 those in Table 
[t_fixedeffects]

).

[^16]: The TOT will be an upper bound if the firms who took up the intervention
 were the ones with most to gain from exporting (although we find no evidence
 of selection into takeup based on observables).

[^17]: Note that the ITT and TOT do not perfectly scale up by the takeup rates
 shown in Table 
[t_takeup]

 since a handful of firms that eventually took up had not yet done so by
 the first follow-up rounds.

[^18]: fn:stress

Of course this measure will still be misleading if owners of treatment firms
 increased their effort per hour worked.
 We were unable to devise survey questions that could accurately capture
 worker effort, but we did ask about job stress measured on a scale of 1
 to 5.
 Reassuringly, when we regress job stress on log output per hour, treatment
 and the interaction of the two (plus round and strata fixed effects), the
 coefficients are neither individually nor jointly significant (F-stat of
 1.40).

[^19]: For the subset of firms that do purchase inputs themselves, we subtract
 the prices of the warp and weft thread inputs from 
$p$

 to make these prices comparable across all firms.

[^20]: This method is similar to estimating the impact of treatment on a standardized
 index of quality (e.g.

Kling et al. (2007)

), but we prefer our method as it produces more conservative estimates in
 our data (i.e.
 higher standard errors).

[^21]: Looms do vary by size but we control for loom sizes through strata fixed
 effects in the analysis below.

[^22]: We abstract from material inputs in the production function since, as discussed
 earlier, intermediaries typically provide the raw materials to firms.

[^23]: More precisely, 
$d\lambda{db}=-(qa)_{\lambda}{b(qa)_{\lambda\lambda}+p_{0}a_{\lambda\lambda}}>0$

 if 
$q_{\lambda\lambda}<0$

 and 
$a_{\lambda\lambda}<0$

, where subscripts denote partial derivatives.

[^24]: Specifically, we require 
$qa$

 is supermodular in (
$\lambda$

,
$\chi_{q}$

) for 
$d\lambda{d\chi_{q}}=-b(qa)_{\lambda\chi_{q}}{b(qa)_{\lambda\lambda}+p_{0}a_{\lambda\lambda}}>0$

 (in addition to the concavity in 
$\lambda$

).

[^25]: Specifically, we require 
$(pa)_{\lambda\chi_{a}}>a_{\chi_{a}}{a_{\lambda}}(pa)_{\lambda\lambda}$

 for both 
$dq^{*}{d\chi_{a}}=q_{\lambda}d\lambda{d\chi_{a}}>0$

 and 
$da^{*}{d\chi_{a}}=a_{\chi_{a}}+a_{\lambda}d\lambda{d\chi_{a}}<0$

 to be satisfied (in addition to the concavity in 
$\lambda$

), where 
$d\lambda{d\chi_{a}}=-b(qa)_{\lambda\chi_{a}-p_{0}a_{\lambda\chi_{a}}}{b(qa)_{\lambda\lambda}+p_{0}a_{\lambda\lambda}}$

.

[^26]: For this reason our framework excludes marketing capital as a factor in
 the production function.
 We believe, consistent with the learning-by-exporting literature, that
 the knowledge generated from matching with foreign buyers is not anticipated
 by the firms and hence should not be reflected in productivity.
 See 
[s_append-tfp]

 for further discussion.

[^27]: fn:adjustmentbias

Specifically, due to this selection, the productivity penalty for making
 a high-specification rug will be larger than the coefficients imply.
 If our experiment induced lower-ability firms to make high-specification
 rugs, the ITT that compares specification-adjusted productivity between
 treatment and control would likely be biased downward (because lower-ability
 firms in treatment would 
"

appear
"

 less productive when using the biased coefficients for the adjustment).

[^28]: As before, we account for correlations across quality metrics by stacking
 our metrics, interacting treatment with each metric and a strata fixed
 effect, and clustering standard errors by firm.

[^29]: We use firm fixed effects here rather than baseline controls so that we
 can visualize the changes between baseline and the follow-up survey rounds
 which would not be possible with baseline controls.

[^30]: fn:timepaths

Analyzing the Joint Sample is complicated here by the different timelines
 across samples.
 Figure 
[f_learning_cdf]

 shows the cdf of total export production across the two samples and Figure

[f_learning_days]

 plots days since first order against export production for each firm.
 As Sample 1 firms started exporting earlier they drive all the variation
 at high values of cumulative exports but not at lower values, leading to
 a spurious discontinuity around this transition.
 Accordingly, Figure 
[f_learning_curves]

 restricts attention to the common support; the range of cumulative exports
 achieved by Sample 2 firms (0-727m
2

).
 Figures 
[f_learning_curves_1]

 and 
[f_learning_curves_2]

 present the two samples separately across their full ranges.
 See Table 
[t_dynamics]

 for a parametric analysis that interacts treatment with round dummies and
 finds most learning occurring by the first follow-up round.

[^31]: fn:learnspeed

The finding that learning about quality occurs quickly is consistent with
 other recent studies.
 In a randomized study of management practices in Indian textile firms,

Bloom et al. (2013)

 find reductions in quality defects after just 10 weeks.
 Likewise, 
(Levitt et al. (2013))

 document a 70 percent decline in defect rates in an automobile manufacturing
 firm just 8 weeks after new production processes were introduced.

[^32]: fn:previous

For example, Hamis learned how to reduce the waviness of rugs by regularly
 adjusting the tenseness of the warp thread from a large furniture distributor
 near Bonn, Germany; and how to ensure consistency of designs over large
 orders by using measuring sticks and folding the rugs over to check for
 symmetry from a small artisan rug shop near Mannheim, Germany.

[^33]: Note that we do not include controls for cumulative production or the takeup
 main effect since both vary only across firms and so are swept out by the
 firm fixed effects (as would any specification controls).

[^34]: There is no evidence that firms achieve higher quality on the talked about
 dimension by reducing effort on other dimensions: there is a positive and
 significant coefficient on the interaction term if quality is regressed
 on 
$Takeup_{i}$

 and 
$Takeup_{i}$

 interacted with a dummy for whether Hamis discussed any quality dimension
 with them (with baseline controls in lieu of firm fixed effects).
 This result is available upon request.

[^35]: We find a coefficient on the interaction of 0.05 (s.e.
 0.04).
 As elsewhere, we include round and strata fixed effects.

[^36]: Specifically, we regress the master artisan quality scores in the quality
 lab on log average monthly hours post-baseline among the control group
 (controlling for baseline quality, baseline hours worked, and round and
 strata fixed effects), and apply the (insignificant) point estimate of
 0.036 to the 5 percent increase in labor hours.

[^37]: A regression among takeup firms of the log price received on log total hours
 of visits by the intermediary and specification controls gives a negative
 and insignificant coefficient of -0.06 (s.e.
 0.06).

[^38]: This is a regression of log profits per hour on the 9 quality metrics recorded
 in our 
"

Step 2
"

 quality lab (as well as specification controls and round and strata fixed
 effects).
 The test that all the quality coefficients are jointly zero is rejected
 at the 

[Table file not found: string".tex]


percent level.
 The regression is reported in Table 
[t_hedonics]

.

[^39]: In the Sample 1 baseline, we did not record the market segment or rug difficulty.
 We replace these missing values with the corresponding values from the
 subsequent survey round.

[^40]: We also note that, 


unanticipated or not, to our knowledge resources associated with generating
 export orders have never been explicitly included in any production function
 estimation.
 For example, de Loecker (2007, 2013) includes an export dummy in the productivi
ty law of motion, rather than as a separate factor of production.

[^41]: See de Loecker (2013) for an extensive discussion of this point.
 To check the stability of the production function coefficients, we ran
 an OLS estimation of 
e_tfp_adjust

 on baseline and endline data, and allowed the 
$\alpha_{l}$

 and 
$\alpha_{k}$

 coefficients to vary by round and treatment/control group; we find no statistic
al difference between the coefficients.
 This suggests that while the level of TFP changed among treatment firms,
 their production function did not.

[^42]: For comparison, the OLS of 
e_tfp_unadjust

 gives 
$\alpha_{l}=$



[Table file not found: string".tex]


(s.e.
 of 

[Table file not found: string".tex]
 \unskip

) and 
$\alpha_{k}=$



[Table file not found: string".tex]


(s.e.
 of 

[Table file not found: string".tex]
 \unskip

).

[^43]: Statistics from Euromonitor International Passport Database, Egypt national
 statistics, UN and OECD.

[^44]: See Ahn et al.
 (2011) to see how exports through intermediaries are classified in this
 setting.

[^45]: In addition to purchasing from producers, some intermediaries employ in-house
 workers.
 For example, throughout the project, Hamis employed workers who worked
 on its premises producing samples as well as orders outside our research
 project.

[^46]: A regression of log unadjusted output per hour on having exported at baseline
 gives a coefficient of 

[Table file not found: string".tex]


 (s.e.


[Table file not found: string".tex]
 \unskip

), while a regression of the mean of all quality metrics on having exported
 at baseline gives a coefficient of 

[Table file not found: string".tex]


 (s.e.


[Table file not found: string".tex]
 \unskip

).

[^47]: ATA's grant expired and in September 2012 it formally ended its involvement
 in this project and closed its Cairo office.
 However, Hamis Carpets agreed to continue participating in the research
 experiment after ATA exited for several reasons.
 First, we sponsored the CEO's visit to the New York International Gift
 Fair in January 2013.


 Second, we provided a quarter of the capital ($7,000) to finance a sample
 order for a new client which was ultimately unsuccessful.
 Third, we provided $500 a month to offset costs of participating in the
 experiment (conducting rug quality surveys, filling out order books etc.).
 Finally, the CEO believes that showing how exporting improves the livelihoods
 of the local population will be good for promoting Fowa's weaving industry.

[^48]: As kasaees rugs are substantially quicker to produce we offered an initial
 packet size of 250m
2

 to Kassaes producers.

[^49]: On a more practical note, it would also have been infeasible to demand that
 Hamis continue to work with a firm that was clearly not able to produce
 at an acceptable standard.
 Hamis' foreign buyers are demanding and would not accept subpar rugs.

[^50]: In cases in which the treatment firms were unable to produce the rugs at
 the quality level required by Hamis Carpets, the research team would reimburse
 Hamis carpets the cost of the materials that were used in the first 110m
2

.
 This happened in only 3 cases, and in none of Duble producing-firms that
 comprise the Joint Sample we use in the paper (see Table 
[t_takeup]

).

[^51]: As noted in the draft, one control firm was incorrectly treated due to an
 error by Hamis.
 In the empirical analysis we make the most conservative assumption and
 keep this firm in the control group.

[^52]: 30 of the 35 Sample 2 treatment firms took the offer up immediately in March
 2013.
 The 2 remaining firms began producing orders for Hamis in May 2014.
 This delay was due to capacity constraints on the side of Hamis.

[^53]: Due to funding constraints, we stopped regular data collection on non-duble
 producers after round 3.
 We did, however, collect data in round 5 (effectively the endline round)
 for the tups and kasaees strata but for budgetary reasons we did not include
 the goublan strata as they had the lowest takeup of all strata.



---

## References