Full text of DOI:10.1162/003355300554962

REPUTATION EFFECTS AND THE LIMITS
OF CONTRACTING: A STUDY OF THE
INDIAN SOFTWARE INDUSTRY*
ABHIJ IT V. BANERJ EE AND ESTHER DUFLO
This paper examines evidence of the role that reputation plays in determining
contractual outcomes. We conduct an empirical analysis of the Indian customized
software industry, using a data set we collected containing detailed information on
230 projects carried out by 125 software firms. We propose a model of the industry
where reputation determines contractual outcomes. The evidence supports the
view that reputation matters. Ex ante contracts, as well as the outcome after ex
post renegotiation, vary with firms’ characteristics plausibly associated with
reputation. This holds after controlling for project, client, and firm characteristics.
I. INTRODUCTION
While the idea that there are severe limits to contracting and
therefore reputation matters is now commonplace in economics,1
little is known about the size of these effects. If these effects are
shown to be large, they would imply a distinct view of the
determinants of trade and growth, and, by implication, of policy: a
view in which legal infrastructure and the availability of reputed
firms could be as important as comparative advantage, human
capital, and physical infrastructure.
This paper attempts to quantitatively assess the importance
of reputation and, by implication, the seriousness of the limits on
contracting, using data gathered from the Indian customized
software industry. India is one of the largest exporters of software
to the United States, but its legal infrastructure is widely seen as
being quite primitive, limiting the scope for contracts. Contracts,
however, play a vital role in the customized software industry
since the structure of the industry invites holdups. The production
process takes time—eight months on average—during which the
* We thank Daron Acemoglu, Philippe Aghion, Susan Athey, Roland Be´nabou,
Mathias Dewatripont, David Genesove, J onathan Guryan, Oliver Hart, Douglas
Miller, Dilip Mookherjee, Ashok Rai, Emmanuel Saez, Andrei Shleifer, J ean Tirole,
Richard Zeckhauser, and two referees for helpful comments, and Paul J oskow for
encouragement and support. We acknowledge financial support from the National
Science Foundation, the Alfred P. Sloan Foundation, the Shultz Fund, and the
J ohn D. and Catherine MacArthur Foundation. We are especially grateful to the
software professionals in India, who patiently answered our many questions, and
in particular to Mr. N. R. Narayana Murthy, from Infosys Technologies Limited,
whose help and insights were crucial throughout this project.
1. For an early expression of this view, see Fama [1980].
2000 by the President and Fellows of Harvard College and the Massachusetts Institute of
Technology.
௠
The Quarterly J ournal of Economics, August 2000
989

990
QUARTERLY J OURNAL OF ECONOMICS
firm commits a substantial number of man-hours. The final
product is typically quite specific to the client, so the client forgoes
going to another firm. The client always keeps the property rights,
and there is little general learning in the process of production.2
The extent to which contracts can protect, however, is limited
by the fact that the desired end-product tends to be complex and
difficult to describe ahead of time in a way that a court, especially
an Indian court, would understand. At the outset, in most cases,
neither the client nor the firm entirely understands what they are
setting out to build. Disputes about the interpretation of the
contract are not uncommon whenever there are cost overruns, i.e.,
costs in excess of what was envisaged at the onset of the project.
Both sides in these disputes accuse the other of failing to adhere to
the letter of the contract, with the client claiming the software
firm (hereafter: the firm) has not delivered what it had contracted
to deliver, while the firm argues that the client has changed his
demands. Given that the description of the product in the contract
is highly incomplete, courts cannot be relied on to adjudicate
correctly between these conflicting claims. In other words, con-
tracts that target the source of the overrun are probably unenforce-
able. This is consistent with the fact that in the Indian industry
we only observe fixed-price contracts (where the firm is supposed
to absorb the entire overrun) and time and materials contracts
(where the client is supposed to pay for all of it).
The basic hypothesis of this paper is that a role for reputation
arises because contracts are limited in this way. We model
reputation for being reliable: reliable firms and clients always pay
for any overrun that they have generated, as long as the other side
does the same. While this is what we often heard (many people
told us ‘‘it was our fault and so we paid for it’’), and it is consistent
with the data (see Section III), we recognize that other, related,
forms of reputation are probably important in practice and that
our empirical work cannot distinguish between them.
If all firms and clients were reliable, there would be no need
for contracts about the sharing of overruns. However, it is usually
impossible to be sure if the other side is reliable, hence the
observation that all projects begin with either a fixed-price
contract or a time and materials contract. While these contracts
only protect one side from unreliable behavior by the other, that
2. In 90 percent of the cases we looked at, the platform and programing tools
were known to the firm before starting the project.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
991
still dominates having no contract at all.3 The choice of which side
to protect depends on which side is more likely to be unreliable,
i.e., on the reputations of the two sides.4 This generates correla-
tions between reputation, the chosen contract, and contractual
outcomes (such as the actual sharing of overrun) that we test in
the empirical section.
The model suggested here is formally developed in Section II.
In Sections III and IV we show that the model and its predictions
are consistent with the observed facts. In particular, we find
strong effects of reputation (measured in several different ways)
on the choice of contracts and the contractual outcomes. Crocker
and Reynolds’ [1993] study of the procurement of airplane engines
by the military is, to our knowledge, the only other paper that
documents the effect of reputation (measured in their case by past
disputes) on the choice of contracts (but not on the sharing of
costs). The results of McMillan and Woodruff [1999] are also
related. Using data from Vietnam and using several measures of
trust including the length of the relationship and information
from third parties, they show that interfirm trade credit is more
likely when the delivering firm trusts its client.
II. A MODEL OF THE SOFTWARE INDUSTRY
The model we propose in this section is an attempt to capture,
in as simple a way as possible, the story of the industry told in the
introductory section. The model adds endogenous contractual
choice to a basic model of reputation formation along the lines of
Diamond [1989].
II.1. Disagreements, Overrun, and Contracts
The client (C) wants the firm (F ) to build a piece of software
that will be worth V to the client (we will assume risk neutrality
3. The ineffectiveness of the Indian court system also limits the extent to
which these contracts protect. The costlier, more corrupt, and slower the courts,
the larger will be the concessions that the plaintiff will be willing to make in an
out-of-court settlement. However, this is not to say that contracts are irrelevant.
The renegotiated outcomes in our sample are indeed correlated with the initial
contract. Anecdotal evidence suggests that some disputes do end up in court (at
least one of the firms in our sample went to court). An Indian legal expert
suggested that unless the plaintiff provides prima facie evidence that the court can
interpret, the case has a chance of getting thrown out, which in this case amounts
to upholding the written contract.
4. For an alternative view of what determines the choice between fixed-cost
and time and materials contracts based on ease of renegotiation, see Bajari and
Tadelis [1999].

992
QUARTERLY J OURNAL OF ECONOMICS
on both sides throughout, so that this is best thought of as a
money payoff and the costs as money costs). In a world where the
client can describe the product it wants perfectly and the firm also
understands this description perfectly, the project should cost y.
We adopt the normalization that y ϭ 0. We assume, however, that
in every real project the actual cost will be positive; i.e., there will
be overrun. One reason for this is that it is very costly for the
client to describe ex ante exactly what it wants, and even if the
description is perfect, it is too costly for the firm to understand it
perfectly. Thus, total overrun will be the sum of overrun caused by
the firm ( yF ), and overrun caused by the client ( yC). We assume
that all overrun is initially paid for by the firm.
The amounts of the overrun, yF and yC, are a result of choices
made, respectively, by the firm and the client. Specifically, we
assume that firms face a choice between a high level yF and a low
level yF. Likewise, the client faces a choice between yC and yC.
Firms can reduce overrun by putting more effort into understand-
ing what the client wants, by better management, and by assign-
ing more able people to the project. Clients can reduce overrun by
working more closely with the firm, and in particular, by being
precise about what they want. Reducing overrun is therefore
costly. The extra costs to the client and the firm from choosing a
low level of overrun are, respectively, BC and BF. Assume that yC Ϫ
yC Ͼ BC and yF Ϫ yF Ͼ BF so that it is always efficient to minimize
overrun.
However, we will assume that both yC and yF are known only
to the firm and its client. Third parties, such as the courts, only
observe total overrun ( yC ϩ yF ).5 Moreover, we restrict all con-
tracts to being linear and additionally require that they do not
involve throwing away any money. In other words, we only
consider contracts where the client pays the firm an amount P ϩ
(1 Ϫ s)( yF ϩ yC), where P is a prespecified fixed payment and s is
the share of the overrun borne by the firm (s ʦ [0,1])—and neither
party makes any payments to third parties. Of particular interest
to us will be two extreme contracts: the contract with s ϭ 1
(corresponding to a fixed-price contract) and the contract with s ϭ
5. In other words, the fact that, say, two people were assigned to the project
for fourteen weeks longer than originally planned is verifiable, but not whether it
took that long because the firm had not understood what it had to do or because the
client had changed his mind about what he wanted.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
993
0 (which corresponds roughly to a time and materials contract).6 It
is key to our analysis that neither of these contracts obviously
dominates the other. The fundamental trade-off comes from the
fact that s is one number that is being used to give incentives to
both parties. A high s gives good incentives to the firm but not to
the client, while a low s contract does the reverse. As a result, it is
typically not possible to implement the first best.7
II.2. Norms, Courts, and Renegotiation
When the first best cannot be achieved by contractual means,
it is possible to improve on the outcome if the behavior of the firms
and the clients is at least partly norm-governed. Specifically,
assume that there are two types of firms and two types of clients.
Of these, one type of firm and one type of client observe a norm of
being reliable. Reliable firms and clients pay for the overrun that
they generate, as long as the other side is reliable, but act to
maximize current profits otherwise. Assume by contrast that
unreliable firms and clients are myopic and always act to maxi-
mize their current profits, although it is possible that this involves
simulating reliability.
When a firm and a client are matched, they do not directly
observe each other ’s types. Rather, on the basis of what it knows of
the client (firm), the firm (client) puts a probability ␪ (␪F ) on the
C
client’s (firm’s) reliability. ␪C and ␪ therefore measure reputation.
F
Our formulation leaves out one of the key ingredients of the
standard reputation model, the type that Tirole [1996] calls
opportunists who choose to be reliable, not out of inner compul-
sions but because it pays in the long run. However, note that in a
reputational equilibrium this type is behaviorally indistinguish-
able from the true reliable type. Therefore, we can accommodate
the possibility that a fraction of the reliable types are actually
reliable by choice as long as we assume that there is some
probability that unreliable behavior will become public and that
in equilibrium no one contracts with those who are known to have
6. The correspondence is not exact because a time and materials contract
typically pays a markup on the realized costs, rather than a fixed payoff. Similar
results hold for that case but the exposition is somewhat more cumbersome.
7. This basic tension is very general. Our restriction to linear contracts, while
vital in the discrete case we have chosen here, can be relaxed if we are prepared to
go to the model where overrun varies continuously: the impossibility of implement-
ing the first best in that case is a consequence of the results in Holmstrom [1982].
More intuitively, the fact that both sides can generate overrun but the courts
cannot observe who was responsible generates a free-rider problem. This is what
drives our model.

994
QUARTERLY J OURNAL OF ECONOMICS
been unreliable in the past (as in Kandori [1992] or Tirole [1996]).
Introducing into our model opportunists who only sometimes
behave reliably complicates the analysis but leaves the main
results unchanged.
In our chosen setting, since the reliable firms and clients are
going to be self-regulated, the function of the contract is to protect
reliable clients against opportunism by unreliable firms and vice
versa. When at least one side turns out to be unreliable, there is a
dispute, and the contract becomes important. However, because
going to court is costly, the contract is likely to be renegotiated. We
assume that in the renegotiated outcome, the firm pays a share
s*(s) of the overrun when there is a dispute. It is natural to
assume that s* is increasing in s. We also assume that s* is
bounded below by s* Ͼ 0 and above by s* Ͻ 1.8 A higher s* and a
lower s* correspond in a natural way to a less effective legal
system.
The timing of the model is as follows: a firm and a client are
matched and decide on a contract; the project is then carried out;
people choose actions that lead to an overrun. At the end of the
project, the overrun is observed, and the two parties decide
whether or not to behave reliably. If either side is unreliable, there
is a dispute, and the overrun is shared in the proportion s*(s)/
(1 Ϫ s*(s)).
II.3. Reputation Formation
There are several mechanisms by which reputation can
evolve. First, in those cases where the firm and the client have
contracted at least once before, the presumption is that both had
behaved reliably so that they both now have a better reputation
with the other.
Second, the age of the firm should be a source of reputation.
We have already noted that in order for there to be a reputational
equilibrium, there must be some probability that unreliable
behavior becomes publicly known in the future. Once that hap-
pens, no one will want to contract with the guilty firm, and it will
probably end up going out of business. This selection process
ensures that older firms (and clients) will typically be more
reliable. This is reinforced by the fact that information is revealed
8. Given that we have assumed risk neutrality, the natural interpretation for
s* is that it is the expected share of the overrun borne by the firm in the event of a
dispute. The actual share will presumably vary according to the exact circum-
stances of the negotiation process, and, on occasion, may turn out to be 1 or 0.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
995
over time, and therefore a firm that has been in the industry a
long time and does not have any black marks against it, is more
likely to be reliable.
Finally, firms may be able to establish a good reputation by
demonstrating that they follow processes that, in principle,
should reduce overrun. Process certification by outside agencies,
such as ISO 9001 certification, is therefore another potential
source of reputation.
II.4. Analysis of the Model
Firm Behavior. Reliable firms have only one choice, to choose
between yF and yF. Unreliable firms have the same choice, but in
addition they can choose whether or not to mimic reliable firms in
the current period. Clients face an exactly symmetrical choice.
The following proposition gives sufficient conditions for there to be
an equilibrium where reliable firms choose low overrun, and
unreliable firms behave unreliably and choose high overrun.
CLAIM 1. The unique equilibrium behavior of reliable firms and
clients is to choose yF ϭ yF and yC ϭ y , as long as (␪ ϩ
C
C
s*(1 Ϫ ␪C))( y Ϫ y ) Ͼ BF and (␪ ϩ (1 Ϫ s*)(1 Ϫ ␪F ))(y Ϫ
F
F
F
C
yC) Ͼ BC. The unique equilibrium behavior of unreliable firms
and clients is to choose yF ϭ yF and yC ϭ yC, subsequently
going into dispute, as long as s*( yF Ϫ yF ) Ͻ BF, (1 Ϫ s*) ·
( yC Ϫ yC) Ͻ BC, and yF and yC are both sufficiently close to 0.
For these conditions to be jointly satisfied, ␪ and ␪ have to
F
C
be high enough, and s* must be sufficiently bounded away from
both 0 and 1. If the first condition fails, i.e., if either ␪ or ␪C is not
F
large enough, it does not pay to be reliable as the other side is very
likely to be unreliable. When s* is very large, it implies that the
contract is very effective in punishing unreliable behavior by firms
and therefore even unreliable firms will behave reliably. The same
holds when s* is very small. In the rest of the formal analysis, we
will assume that these conditions hold.
The Optimal Contract and the Sharing of Overrun. The firm
proposing the contract in our model knows it own type, and as
result the proposed contract can be used as a signaling device. A
firm that plans to be unreliable will prefer a contract where it pays
very little of the overrun. Therefore, by choosing to absorb most of
the overrun, a firm may be able to signal that it is reliable. Given
that we are in a signaling environment, we expect that there will
be many equilibria. However, all such equilibria will involve

996
QUARTERLY J OURNAL OF ECONOMICS
pooling because in any separating equilibrium no one will contract
with unreliable firms.
Among the set of pooling equilibria, we focus on contractual
outcomes where the total joint surplus of a firm and a client who
are both reliable is maximized. This is always a Bayesian-Nash
equilibrium (sustained by the belief that only opportunists devi-
ate). The fact that it is also Pareto optimal from the point of view
of the reliable types makes it an obvious focal outcome.
This expression for joint surplus is
W(s,␪C,␪F ) ϭ V Ϫ ␪FyC Ϫ (1 Ϫ ␪F )(1 Ϫ s*(s))( yC ϩ yF )
Ϫ ␪CyF Ϫ (1 Ϫ ␪C)s*(s)( yF ϩ yC).
The third term in this expression gives the total surplus that
is lost because in a pooling equilibrium a reliable client must allow
for the possibility that the firm is unreliable, while the fifth term
is the surplus that is lost because the firm must allow for the
possibility that the client is unreliable.
From the linearity of this expression in s*, it immediately
follows that the optimal contract is always either a fixed-price or a
time and materials contract (s is either 0 or 1). For any fixed level
of client reputation, the more reputed the firm, the more likely it is
that it is a time and materials contract.9
The expected share of the overrun paid by the average firm
with reputation ␪ that works for a client of reputation ␪C, is ␪ ␪
F C
F
[y /(y ϩ y )] ϩ (1 Ϫ ␪ ␪C)s*(s(␪F,␪C)). Increasing ␪ has two ef-
F
F
F
F
C
fects on the expected share. First there is an effect through the
choice of the contract. Second, keeping the choice of the contract
fixed, there is a mechanical effect deriving from the fact that more
reputed firms are more likely to behave reliably and keep their
overrun under control. As a result, they end up paying a smaller
share of the overrun as long as s* Ͼ [yF /( yF ϩ yC)], which is likely
for firms with low reputations since they face fixed-price con-
tracts. However, note that for very reputed firms, as it is likely
that s* Ͻ [yF /( yF ϩ yC)], the net effect of an increase in reputation
can be to increase their share of the overrun.
9. The reader may feel that the result that the contracts are always at one or
the other extreme is driven by the assumption of risk neutrality. This is partly true
in the sense that if the two parties are sufficiently risk-averse a less extreme
contract will be chosen. However, note also that s* is always strictly between zero
and one; the effective contract is never very extreme. Therefore, it may well be the
case that the initial choice of an extreme contract is consistent with optimal risk
sharing.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
997
The mean overrun generated by a firm of reputation ␪ and a
F
client of reputation ␪ is ␪Fy ϩ (1 Ϫ ␪F )y ϩ ␪Cy ϩ (1 Ϫ ␪C)y .
C
F
F
C
C
This is clearly decreasing in ␪F. Finally, the share of total overrun
caused by the firm is ␪ ␪C[y /(y ϩ y )] ϩ ␪F(1 Ϫ ␪C)[y /(y ϩ y )] ϩ
F
F
F
C
F
F
C
␪C(1 Ϫ ␪F ) [yF /( yF ϩ yC)] ϩ (1 Ϫ ␪F )(1 Ϫ ␪C)[yF /( yF ϩ yC)]. Simple
calculations show that this expression is decreasing in ␪F.
So far, our results describe the effect of increasing ␪F, keeping
␪C fixed. However, our data do not contain measures of the client’s
reputation. The same results, however, hold without conditioning
for the client’s reputation under two scenarios. The first is one
where most clients have similar reputations. This is not implau-
sible because most clients are well-established companies. Alter-
natively, we could assume that there is random matching. This
may be plausible in at least some cases because clients have quite
specific demands. Finding an Indian firm with the right specializa-
tion that can work on the required schedule may not always be
easy, forcing the client to choose the one firm that, by chance,
happens to be available at that time.10 Under either of these
scenarios we have
CLAIM 2. Firms with a better reputation are more likely to have
time and materials contracts. In most cases they also pay for
a smaller share of the overrun. There is less firm-generated
overrun and less total overrun in contracts involving more
reputable firms.
II.5. Discussion
Our model yields a number of clear-cut results, but only
under rather stringent conditions. One key assumption was that
all firms do similar projects. If firms in our model could choose
their projects, low reputation firms would tend to gravitate
10. The random matching assumption is clearly indefensible in the case of
repeat contracts. The fact that the client and the firm have agreed to work together
again signals mutual respect. Compared with the initial match between them,
both ␪ and ␪ should therefore be higher. The net effect is therefore potentially
C
ambiguous. HFowever, given that there are many more start-ups among firms than
among clients, it is plausible the firm’s reputation will improve by more than the
client’s reputation. In this case, repeat contracts will be more likely to be time and
materials contracts, and the firm’s share of the overrun in a repeated contract
should be lower. For the case of first time matches, in a previous version of this
paper we formalized an argument showing that, ceteris paribus, firms with low
reputations will prefer to work with highly reputed clients. This is because they
are particularly vulnerable to opportunistic behavior by the client. This kind of
nonrandom matching reinforces our result that low reputation firms get fixed-
price contracts and end up with a higher share of the overrun. However, it also
weakens the relation between total overrun and the firm’s reputation.

998
QUARTERLY J OURNAL OF ECONOMICS
toward projects with low yC even at the cost of lower returns. This
is because they have the most to lose if the client misbehaves.11
This suggests that low reputation firms will tend to choose
projects which are simple and well understood where the client
does not have to do very much work to make clear what he wants.
Firms that switch to a low yC project should be more willing to
accept a fixed-cost contract. Moreover, the switch to low yC will
tend to raise the fraction of the overrun generated by the firm, and
therefore the firm will typically pay a larger part of the overrun.
All of these effects go in the direction of our previous results. Total
overrun may, however, no longer fall with reputation since low
reputation firms choose low overrun projects.
Our model is also limited by the fact that, by assumption,
contracts have no direct incentive effects. If the set of options
faced by the firms and clients were less discrete, one would
imagine that contracts would also have incentive effects. A high
value of s would then have the additional effect of discouraging
firms from generating very high levels of overrun. An incentive
effect of this kind reinforces the correlation between low reputa-
tion firms and fixed-price contracts. However, it also means that
low reputation firms do not necessarily generate more overrun
than their more reputable counterparts. This, in turn, weakens
the correlation between reputation and total overrun. It should
also weaken the relation between reputation and the share of
overrun that is paid by the firm. However, one would still expect
this last correlation to be negative, as low reputation firms have
fixed-price contracts and those who have fixed-price contracts pay
for more of the overrun.
A generalized version of our basic model that allows for both
project choice and the incentive effect of contracts would have the
following predictions: 1) low reputation firms will choose projects
that are simple and easy to define; 2) low reputation firms are
more likely to face fixed-price contracts, and on average pay for a
higher share of the overrun; and, 3) neither total overrun nor the
overrun generated by the firm is necessarily correlated with its
reputation.
The combination of the last two predictions is what distin-
guishes a model, like ours, that emphasizes incentive problems
and reputation effects: low reputation firms in our model will face
harsher terms even if their actual performance (measured by
11. This is formally shown in a previous version of the paper.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
999
overrun) is indistinguishable from that of high reputation firms.12
This is because what matters in this type of model is not what
firms actually do, but rather, what they could do.
By contrast, theories where firms are paid based on perfor-
mance, but not because of contracting problems—standard mar-
ginal product theories, for example—would predict a correlation
between reputation and the share of overrun paid by the firm, only
if there was also a correlation between reputation and overrun
caused by the firm (which, in turn, suggests a correlation between
reputation and total overrun).13 Theories that give a central role
to bargaining between the firm and the client, for example, can
generate a correlation between reputation and the share paid by
the firm if bargaining power is correlated with reputation, but
would tend to predict the wrong correlation between contract type
and reputation. Intuitively, a firm that has a lot of bargaining
power will value having a fixed-cost contract less than a firm that
is in weak bargaining position. Finally, risk-sharing-based theo-
ries would tend to predict the wrong correlation between reputa-
tion and the share of the overrun paid by the firm. Less reputed
firms are typically smaller and have fewer assets, and should bear
less of the risk (overrun).
III. DESCRIPTIVE EVIDENCE
In this section and the following, we document that both the
central assumptions and main implications of our model are
consistent with what we observe in the industry. The data set used
here was collected by interviews with the CEOs of 125 software
firms in the three major centers of software production in India
(Bangalore, Pune, and Hyderabad) in the winter of 1997–1998.
We collected general information about each firm, and specific
data about the last two projects completed by the firm. More
information about the data collection process is presented in the
Appendix.
12. There are of course other models within the general category of incentive/
reputation-based models. For example, a model where the reputation is for quality
of work rather than for reliability may have rather similar predictions.
13. One example of this class of theories would be one that says low
reputation firms generate more overrun and therefore pay for them. An alternative
would be the argument that low reputation firms set low estimates either
deliberately or out of inexperience and end up paying more as a result.

1000
QUARTERLY J OURNAL OF ECONOMICS
III.1. Projects, Estimates, and Overruns
We focus in this paper on contracts for customized software
projects. A project is a well-defined notion in the industry. It
corresponds to a set of functions that the software must execute.
These functions are (often loosely) described in the Request for
Proposal (RFP) sent by the client when they first look for a firm.
Most projects in the sample (75 percent) were for overseas clients.
Common projects include the development of a customized appli-
cation specific to the client, customization of a package, porting or
reengineering of an existing solution, and testing of existing
software. Seven percent of the projects in the sample were Y2K
projects, and another six percent were other simple projects: data
manipulation, web pages, and CAD projects.
In response to the client’s RFP, several firms submit propos-
als. Then the client chooses one of these firms. Each proposal
contains contractual terms and an estimate. Our understanding,
based on descriptions in standard software management text-
books (e.g., Pressman [1997]) and conversations with people in the
industry, is that the estimate is supposed to be the firm’s best
guess as to how many hours will be needed to complete the project,
assuming that the firm’s current understanding of the project is
correct and that the firm adheres to its own productivity norms.14
The firm and the client understand that both these assumptions
may be false. The project may take more effort than estimated
either because the firm took the wrong direction on the project, or
the client changed his demands, or because productivity in the
firm was lower than it would normally be. This uncertainty is
taken into account in the terms of the contract, but the estimate is
still a useful benchmark.
We asked the firm what the initial estimate was for each
project. Using a timeline representing the project (which we drew
on a piece of paper as we conducted the interview), we then asked
them to identify all the stages in the project when the estimate
was revised and by how much (more details about the process are
given in the Data Appendix). Overrun is defined to be the
difference between the actual effort involved in completing the
project and the estimate. Several different questions were asked
about the level of the overrun, and we made sure that the
14. The software industry measures everything in the unit of man-hours.
This reflects the fact that labor is the key input and all other inputs (computers,
space, and phone time) are more or less linearly related to the number of
employees.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
1001
respondent gave consistent answers to all the questions, probing
him if there was an inconsistency. We are confident that the figure
we obtained for the overrun is a reliable measure of the difference
between actual cost and the estimate. It is, of course, possible that
the estimate itself was manipulated. A high estimate would favor
an unreliable firm and a low estimate an unreliable client, and
moreover, the estimate could be used for signaling at the precon-
tracting stage. However, the fact that the client gets several
competing estimates and the fact that there are standard methods
and computer packages (used in 69 percent of cases) for comput-
ing the estimate make it much less likely that a firm would get
away with substantially manipulating this number (the evidence
in Section IV on the level of overrun also gives us some confidence
in the way we use these numbers).
Descriptive statistics are presented in Table I. Overruns are
frequent (75 percent of the cases), large (23 percent of the
estimate on average), and variable (the standard deviation is 30
percent). We asked the firms to report the reason for the overrun
by asking them what exactly happened in as neutral a way as
possible. Firms attribute, on average, 45 percent of the overrun to
changes explicitly required by the client, a further 12.8 percent to
delay in the client’s response at critical points, 19.5 percent to
their own misunderstanding of the specifications of the project,
and 9.1 percent to internal problems in the firm (for example, the
loss of the project’s manager). According to this, the client is
responsible for 58 percent of the overrun on average (changes and
delays in providing critical inputs), and the firm for 29 percent
(misunderstandings and internal problems).
III.2. Clients and Contracts
Clients fall into two categories, external and internal clients.
An internal client can be the firm’s mother company (an American
firm or an Indian group), or a company with whom the firm has a
very long-term and open-ended relationship. For example, some
software firms have established ‘‘Off-shore Software Development
Centers’’ (OSDC), for their foreign clients whereby the firm
dedicates a part of its office and manpower to that client. Even in
the case of internal clients, the work is defined in terms of projects
and is governed by a contract. However, the scope for unreliable
behavior in this type of relationship is presumably much more
limited, since both parties share the control rights. We will
therefore view this kind of relationship as a substitute for

1002
QUARTERLY J OURNAL OF ECONOMICS
TABLE I
DESCRIPTIVE STATISTICS
Internal
projects
External projects
Standard
Mean Median deviation Mean
(1)
(2)
(3)
(4)
Number of observations
167
57
PANEL A: PROJ ECT CHARACTERISTICS
Project size (man-months)
Application area is familiar to the firm
Platform is familiar to the firm
Programming tools are familiar to the firm 0.87
Y2K,CAD, Web pages
75
0.68
0.90
27
1
1
1
0
141
207
0.69
0.93
0.84
0.07
0.47
0.30
0.34
0.35
0.14
PANEL B: REPUTATION
Founded 1993 or later
ISO certification
Worked with the client in the past
PANEL C: SOFTWARE FIRMS AND
CLIENTS
0.43
0.19
0.41
0
0
0
0.50
0.39
0.49
0.57
0.14
0.83
Number of employees in the software firm
Client is a Fortune 500 company or an
Indian group
147
60
279
153
0.55
0.22
1
0
0.50
0.42
0.28
0.01
Client is an Indian firm
PANEL D: OVERRUNS
Project cost Ͼ estimate (overrun)
Project cost Ͻ estimate (underrun)
Total overrun (percent of project cost)
Percent of overrun due to the client:
—Changes
0.75
0.02
23
58
45
1
0
15
50
50
0
0.43
0.15
30
43
43
0.70
0.10
27
58
47
—Client’s delay
13
26
11
Percent of overrun due to the firm:
—Ambiguity in specifications
—Internal difficulties
29
20
9
0
0
0
43
34
25
11
20
Ϫ9
PANEL E: CONTRACTS (proportion)
Fixed-cost contracts
Mixed contracts
Time and material contracts
PANEL F: SHARE OF OVERRUN PAID
BY THE FIRM (percent)
All contracts
0.58
0.26
0.15
1
0
0
0.49
0.44
0.36
0.23
0.07
0.67
57
66
28
86
61
77
100
0
100
100
45
44
39
20
45
8
Fixed-cost contracts
Time and material contracts
Fixed-cost contract, firm responsible
Fixed-cost contracts, client responsible
33 no obs.
46 45

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
1003
reputation. Except for a few very large internal projects, internal
and external projects are similar in nature.15 Firms with internal
projects are on average smaller and more recently established
than other firms. Most firms who work for internal clients also do
some work for external clients. We will use this feature to contrast
the features of internal and external contracts.
Consistent with the predictions of our model, we observe only
three types of contracts. Time and material contracts are the least
frequent (15 percent). Fixed-cost contracts are the most frequent
(58 percent). Mixed contracts divide the project into a specification
phase (which is the first and most open-ended part of the project,
where the client’s needs are analyzed to design the broad architec-
ture of the piece of software), and an execution phase (design,
coding, and testing). A separate contract (time and materials or
fixed-price) is signed for each part. The most frequent configura-
tion is a time and materials contract for the specifications, and a
fixed-price contract for the execution.
As assumed in our model, contracts are frequently renegoti-
ated, but do indeed matter for the final outcome. This is shown in
Panel F of Table I. Firms bear on average 57 percent of the
overrun. When they have a fixed-price contract, they bear on
average 66 percent of the overrun (the median is 100 percent).
When they have a time and materials contract, they bear on
average 26 percent of it (the median is 0 percent). The data are
consistent with our premise that some firms tend to pay for their
mistakes no matter what the contract was. Even with fixed-price
contracts, firms pay more (86 percent) of the overrun when they
are fully responsible for it than when the client is fully responsible
for it (61 percent). The ordering is the same for time and materials
contracts.
III.3. Reputation
As discussed in the theory section, we consider three mea-
sures of reputation: the age of the firm, the fact that the firm has
previously worked with the client (41 percent of the contracts),
and ISO certification (19 percent of the contracts). In addition, we
contrast internal and external contracts among firms that do
some internal contracts.
15. The difference in mean effort (207 man-months for internal projects, 75
man-months for external projects) is driven by a few very large internal contracts.
The median efforts are similar in both cases (25 and 27 man-months, respectively).

1004
QUARTERLY J OURNAL OF ECONOMICS
FIGURE I
Proportion of Fixed Cost Contracts
Preliminary evidence suggests that the principal empirical
implications of our model are not rejected by the data. In Figure I
we plot the proportion of fixed-cost contracts as a function of the
age of the firm. In Figure II we plot the average share of overrun
that is paid by the firm. Both the proportion of fixed cost contracts
and the firm’s share of the overrun strongly decrease with the age
of the firm, in particular in the first five years (which is consistent
with the prediction that the marginal effect of reputation declines
as reputation increases). Table II shows the proportion of fixed-
cost contracts, the firm’s share of the overrun, and other project
characteristics for each type of firm, and the difference between
low and high reputation firms. Standard errors of the means and
the difference in means are presented in parentheses. In columns
(1) to (3) we show the contrast between young firms (created in
1994 or after) and old firms (created in 1993 or before). Young
firms are significantly more likely to have fixed-price contracts
(the probability is 26 percent higher). They also, on average, bear
substantially more of the overrun (19 percent more). These
differences are significant at the 95 percent level of confidence.
ISO-certified firms (columns (4) to (6)), by contrast, are no less
likely to get fixed-price contracts nor to pay less overrun than
other firms. Firms engaged in a repeated relationship with their

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
1005
FIGURE II
Share of Overrun Paid for by the Firm
client are 9 percent less likely than other firms to have fixed-price
contracts (columns (7) to (9)), although the difference is not
significant, and they pay significantly less of the overrun (20
percent less). Finally, firms that have internal contracts pay for
more of the overrun when they deal with external clients than
when they deal with internal clients. Almost half of their external
contracts are fixed-price contracts, whereas only 23 percent of the
internal contracts are fixed-price contracts. Firms pay a much
smaller share of the overrun (20 percent instead of 47 percent) in
internal contracts than in external contracts.
This set of contrasts among the contracts and contractual
outcomes of the different types of firms is not explained by obvious
differences in firms’ behavior or abilities, at least to the extent
that these would be reflected in different overruns or different
probabilities of delay.16 In Table II we present the total overrun of
the project, the total overrun due to the firm, and the proportion of
projects that were delivered with a delay. None of the differences
between high and low reputation firms are significant. If any-
16. Delay is clearly an important index of the firm’s performance. For simple
projects where quality can be controlled relatively easily, clients are primarily
concerned about delays (in 54 percent of the cases the deadline was extremely
important or critical to the client). Delay is useful for us because it provides a check
on the firm’s performance that does not rely on the estimate which, as we have
noted, might be subject to some manipulation.

1006
QUARTERLY J OURNAL OF ECONOMICS

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
1007
thing, total overruns and the probability of delay are somewhat
smaller for young firms than for old firms, for non-ISO firms than
for ISO firms, and for external contracts than for internal
contracts.
We interpret these results as showing that reputation (except
for ISO certification) does influence the way the overruns are
shared between the client and the firm. To confirm this interpreta-
tion, we now turn to regressions estimates, which allow us to
control for project, firm, and client characteristics which may have
independent effects on both the contract and the outcome and may
be correlated with our measures of reputation.
IV. EVIDENCE FROM REGRESSIONS
IV.1. Choice of Contract
The theory suggests that there is a threshold, such that when
the reputation of a firm crosses it, the optimal contract switches
from time and materials to fixed-price contracts. In practice, this
threshold should vary by firm, client, and project characteristics,
and perhaps also in response to other environmental features.
Therefore, the natural model to estimate is a discrete choice
model. In our setting there are three kinds of contracts that can be
ordered. Fixed-cost contracts give the most incentives to the firms,
time and materials contracts give the most incentives to the
clients, and mixed contracts are in between. We therefore assign a
‘‘1’’ to a fixed-price contract, a ‘‘2’’ to a mixed contract, and a ‘‘3’’ to a
time and materials contract, and estimate an ordered probit
model for the choice of contract.
Empirically, potential determinants of contract choice include
not only reputation characteristics (a dummy for whether the firm
was created after 1993, a dummy for whether the firm has worked
with the client in the past, and a dummy for whether the firm is
ISO certified) but also other variables potentially correlated with
reputation, which may affect the choice of contract.
First, if firms and clients were risk-averse (in contrast to
what we assume in the theory section), contracts would reflect
risk-sharing motives. The contract may then depend on the
client’s and firm’s size (as a proxy for their risk aversion). The
inherent uncertainty of the project (which should be correlated
with estimated project size, the complexity of the project, the
degree to which the firm and the client are familiar with the

1008
QUARTERLY J OURNAL OF ECONOMICS
project, etc.) should also influence the choice of contract. Since the
client is generally much bigger than the firm, we might expect
that characteristics which make the project riskier will also make
it more likely that the chosen contract is a time and materials or a
mixed contract.
Second, Crocker and Reynolds [1993] argue that fixed-price
contracts are more likely when it is easier to draft an exhaustive
list of requirements (a complete contract). Their empirical analy-
sis confirms these predictions. Contracts tend to be fixed-price if
the engine is well-known or the production cycle is short. In our
context, this would also suggest that the contract is more likely to
be a time and materials contract if the project is large and if the
area of application or the platform is not familiar to the firm. We
would also expect that extremely well-defined projects (such as
CAD or Y2K projects) are more likely to have fixed-price contracts.
To summarize, denote by Cic the type of contract obtained by
firm i in contract c, by Ric the vector of reputation variables, by Xic
the project characteristics, by Zic the firm’s characteristics, and by
Mic the client’s characteristics. We estimate the following model by
ordered probit:17
(1)
Cic ϭ ␣Ric ϩ ␤Xic ϩ ␥Zic ϩ ␦Mic ϩ vi ϩ wic.
Results are presented in Table III, columns (1) and (2). A
negative coefficient on a variable indicates that the contract is
moving away from time and materials toward fixed-price. Control-
ling for client and firm size, the client’s country, and project
characteristics, young firms are significantly more likely to have
fixed-price contracts. Firms that work with a repeated client are
more likely to have a time and materials contract, but the
difference is not significant. ISO-certified firms are more likely to
have a fixed-price contract, although this coefficient is not signifi-
cant. These results confirm the descriptive statistics presented in
Table II.18 In column (2) we focus on firms that have internal
clients. When these firms work with an internal client, they are
significantly more likely to have a time and materials contract
than when they work with an external client.
Interestingly, project characteristics do not seem to influence
17. The standard errors are corrected for firm-level grouping.
18. We have also interacted the fact of working for a repeat client with the age
of the firm. Working with a repeat client makes it more likely that they have a time
and materials contract, and this is not true for old firms. This suggests that old
firms may have a strong enough reputation in the market to make specific bilateral
relationships relatively unimportant.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
1009
TABLE III
REGRESSION RESULTS: CHOICE OF CONTRACTS AND SHARE OF OVERRUN
PAID BY THE FIRM
Share of overrun paid by the firm
Unconditional Conditional
Random Fixed
Choice of
contract
Ordered
probit
Random
effect
Fixed
effect
effect
effect
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Repu tation
Young firm
Ϫ0.69*
(0.25)
0.22
(0.24)
Ϫ0.27
(0.32)
15*
(8.5)
Ϫ17*
(8.8)
17
9.0
(8.6)
Ϫ15*
(8.7)
16
Repeated contract
ISO-certified firm
Internal project
Ϫ20
(16)
Ϫ19
(17)
(13)
(13)
0.87*
(0.31)
Ϫ25*
(11)
Ϫ64*
(26)
Con tr act
Fixed-cost contract
13
12
(9.4)
(25)
Time and material
contract
Ϫ12
15
(13)
(34)
Fir m an d clien t
ch ar acter istics
Number of employees
(/100)
0.44
(0.98)
0.15
Ϫ0.44 Ϫ4.1 Ϫ2.4
(1.1) (7.7) (5.0)
Ϫ0.18 Ϫ17 Ϫ13 Ϫ16 Ϫ40
Ϫ4.8
(4.5)
16
Client is big
Ϫ16
(16)
Ϫ45*
(21)
(0.22) (0.30) (8.5) (10) (16) (32)
Ϫ0.43* Ϫ0.76 13 23 Ϫ46* 14
(0.27) (0.63) (9.5) (18) (20) (41)
(8.3)
9.3
(9.5)
Client is Indian
P r oject ch ar acter-
istics
Estimated project
size (man-
months/10)
0.01
(0.017) (0.011) (7.7) (.38) (1.9) (2.6)
0.08 0.09 Ϫ9.0 Ϫ4.0 18.9 Ϫ1.5
(0.25) (.027) (9.4) (12) (21) (28)
0.60 0.20 26 1.9 48 61
(0.48) (0.59) (20)
Ϫ0.13 0.15 19
(0.32) (0.35) (17)
0.00 Ϫ0.12 0.35 Ϫ2.4 2.1
Ϫ0.13
(0.76)
Ϫ6.3
(9.2)
29
(19)
15
(16)
Ϫ2.1
(2.0)
16
(22)
49
(36)
Ϫ41
(39)
Area is familiar
Platform is familiar
(12) (35) (74)
Ϫ4.8 Ϫ47
(25) (36)
Y2K, data manipu-
lation, etc.
Standard errors (corrected for clustering at the firm level in columns (1) and (2) are in parentheses.
An asterisk indicates that the coefficient is significant at the 90 percent level of confidence.
All regressions include the following additional variables: number of employees squared, project size
squared, and an indicator for whether tools are familiar to the firm. Sample in columns (1), (3), (5), (7), and (8)
include all external projects (167 projects in 92 firms). Sample in columns (2), (4), and (6) include all projects of
firms that do some internal work (88 projects in 58 firms).

1010
QUARTERLY J OURNAL OF ECONOMICS
the choice of contract. Firm and client size have no impact either.
There is no strong evidence that risk-sharing or the difficulty in
drafting the contract (as in Crocker and Reynolds [1993]) plays an
important role in determining the contract.19
IV.2. Share of the Overrun Paid by the Firm
A quantitative measure of the importance of reputation is
obtained by estimating the effect of reputation on the share of the
overrun that is actually paid by the firm. Our model predicts that
reputation influences the share of the overrun paid by the firm
both without conditioning on the contract, and potentially also
after conditioning on the contract. We therefore run both regres-
sions. Clearly, when we do not condition for the contract, we
should control for all the characteristics that affect contractual
choice. Once we condition on the contract, realized overrun and
share paid by the firm are jointly determined. We focus on the
estimation of a reduced form, as we do not have instruments for
unpredicted overrun that would be necessary to estimate a
structural model.
We therefore estimate the model using a random-effects
specification model:
(2)
S
ic ϭ ␣Ric ϩ ␤Xic ϩ ␥Zic ϩ ␦Mic ϩ ␭Cic ϩ vi ϩ wic,
and its equivalent excluding Cic.
The dependent variable of interest is measured only when
there is an overrun.20 We have also estimated a Heckman [1979]
sample selection model, excluding the vector of project char-
acteristics Xic in equation (2). This exclusion is justified, as we
have seen that they do not predict the type of contract. We do not
19. This argument assumes that the project characteristics that we include in
the regression are correlated with project risk and the difficulty of drafting a
complete contract. We present later evidence that overrun is correlated with these
characteristics. We also found (but to save space do not report) that the effort put
into drafting the proposal (a measure of the difficulty to draft the contract) is
correlated with these characteristics.
20. In cases where there is a 0 or negative overrun, we could impute the
sharing of the overrun predicted by the contract since in such cases the contract is
never revised (in fact, 0 overrun often means a very small positive or negative
overrun). The results are similar if we use this alternative variable. However, as
we are interested in how large overruns are shared, it would be misleading to
include the cases where there was very little overrun, because what happened in
these cases does not necessarily tell us what would have happened if the overrun
had been large.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
1011
report the results, which are very similar to the random-effects
specification.21
Results are presented in columns (3), (4), and (7) of Table III.
Young firms pay 15 percent more of the overrun than old firms.
Firms that work with a repeated client pay 17 percent less of the
overrun than other firms. Both of these numbers are significant at
the 90 percent level. ISO firms tend to pay more of the overrun
than non-ISO firms. Firms pay 20 percent less of the overrun
when they do internal projects. Bigger firms, firms that work with
a foreign client, and firms that work for a big client (a Fortune 500
company, or an Indian group) pay less of the overrun. Conditional
on the type of contract (column (7)), young firms pay more of the
overrun (9 percent), but the difference is not significant. Firms
that work with the client for the second time pay more of the
overrun. The difference is as large as in the unconditional
specification and remains significant. The fact that the age effect
is small once we condition on the type of contract is not surprising.
Once we control for contract type, the effect of age in our basic
model comes from the fact that young firms generate more
overrun (this is what we call the mechanical effect). In the more
general model suggested in subsection II.5, the correlation be-
tween age and firm-generated overrun disappears. As we will see,
the data confirm the absence of such a correlation. The strong
effect of repeat contracts, even after we control for the type of
contract, is probably due to an effect that our model does not
capture: in the real world, clients that have had a good experience
with a firm are more like to behave reliably since they want to
retain the firm. In other words, reliability may be endogenous.
We use the fact that we have two observations for most firms
to estimate a fixed-effects model. Fixed-effects wipe out both the
effect of the age of the firm and the effect of whether it is an ISO
firm, thus allowing us to look more closely at the effect of being in
a repeated contract. Specifically, it allows us to check whether this
effect is spuriously generated by the fact that better firms pay less
of the overrun and are more likely to work again with the same
client. It also gives us better estimates of the effect of internal
contracts and other characteristics. Columns (5) and (8) present
results for repeated contracts. Not surprisingly, the fixed-effects
21. We have also experimented with a double-sided tobit model, which takes
into account the fact that the share of the overrun paid by the firm is bounded
between 0 and 100 percent. The results are qualitatively similar; hence, we focus
on the linear model.

1012
QUARTERLY J OURNAL OF ECONOMICS
specification is much noisier than the random-effects specifica-
tion. However, the point estimate of the effect of working with a
repeated client is practically unaffected (Ϫ20 percent instead of
Ϫ17 percent).22 The effect of working with an internal client on the
share of the overrun, shown in column (6), is bigger in the
fixed-effects specification than in the random-effects specification.
IV.3. Overruns and Project Characteristics
An alternative interpretation of our results could be that
some characteristic of the firm, unobserved by us, but known to
the client (say, the quality of the firm) is correlated both with the
choice of contract (e.g., good firms get time and materials con-
tracts) and with our measures of reputation. However, as we have
discussed in the theory section, this class of explanations would
imply a correlation between the performance of the firm and its
reputation. In contrast, our model, in its more general form,
suggests that there can be a correlation between the contractual
outcomes and the reputation of the firm, even in the absence of
any link between the performance of the firm and its reputation.
Therefore, we estimate equations similar to equation (2) but
with a measure of firm performance—such as total overrun (as a
percentage of the initial estimate), overrun generated by the firm,
and delay as the dependent variables. To save space, we report
only random-effects estimates. Fixed-effects point estimates are
similar.
Results are presented in columns (1) to (5) of Table IV.
Neither client overrun nor overrun due to the firm is correlated
with the reputation of the firm. They are not correlated with client
and firm characteristics either, but are correlated with some of the
project characteristics. Shorter projects have smaller overruns in
percentage terms, as do projects where the platform is familiar to
the firm and simple projects such as Y2K or CAD projects. The
coefficients on reputation as well as those corresponding to firm,
client, and project characteristics are similar irrespective of
whether or not we control for the contract. Interestingly, in the
cross section, total overrun (but not the overrun generated by the
firm) seems to be bigger in fixed-cost contracts. If, perhaps as
suggested by Crocker and Reynolds [1993], time and materials
contracts were associated with less effort being put into the
22. The only coefficients affected are those of working with an Indian client
and of doing Y2K projects, which become negative.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
1013

1014
QUARTERLY J OURNAL OF ECONOMICS
precontracting stage, one would have expected the opposite to
hold. However, the association disappears in the fixed-effects
regression. We also find that project delay is not systematically
associated with reputation.
Finally, we examine whether the choice of projects is influ-
enced by a firm’s reputation. Once we control for size, young firms
do not tend to do shorter projects (column (6)). However, consis-
tent with our model, they do more Y2K projects (column (7)), and
their projects are, according to their own assessment, simpler
(column (8)).23
V. CONCLUSION
We set out in this paper to look for evidence that reputation
plays an important role in determining contractual outcomes. We
argue that the evidence strongly supports this view, although
given that we do not directly observe people making use of
reputation, and that there are important unobservables that
might affect our estimates, some doubts clearly remain.
The conclusion that reputation matters is of course impor-
tant, for it gives support to a range of theories that are based on
limitations of contracting. Moreover, it might suggest part of the
explanation as to why the Indian software industry is not much
larger (Indian software exports were only worth 3.4 percent of the
1995 worldwide outsourcing business) despite its obvious labor-
cost advantage24 and the fact that it is a very labor-intensive
industry. Reputation at the firm level is one possible reason. Most
Indian firms are simply not trusted enough to be given important
contracts. While our evidence cannot directly substantiate this
view, the fact that reputation is important within the Indian
industry suggests that it may also be important when an Ameri-
can client is deciding whether to go to a firm in India or to one in
the United States. The industry clearly recognizes this. When we
surveyed the firms, the association of software exporters and the
government were actively supporting the firms to acquire ISO
23. We asked the firms to provide all the metrics that were relevant to the
project: line of codes, number of modules, of screens, etc. After this we asked them
the following: taking everything into account, how would you rate the complexity
of the project (on a scale of 1 to 5)? The dependent variable in column (8) is this
subjective measure.
24. The United States imports a very large number of Indian software
professionals for short-term assignments at a cost of more than twice what they
would earn in India.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
1015
certification as a way of improving their reputation. As we show,
and as they have come to realize, ISO certification is not an
effective source of reputation. The association of software produc-
ers has now switched to recommending a competing standard
(called SEI), specific to the software industry. Our results also
suggest that the industry might benefit from the establishment of
a credible system for rating firms modeled on credit rating
systems, by making it possible for the market to efficiently
aggregate all that is known about each firm.
APPENDIX: DATA COLLECTION
The data set used in this paper was collected by one of the
authors during the period of November 15, 1997, to February 15,
1998. The survey instrument was developed after extensive
conversations with members of the software industry in Banga-
lore during the preceding summer, and pretested on ten firms
before the actual survey began. The sample frame was the list of
firms belonging to the government-exporting organization (the
Software Technology Park). Almost all exporting firms belong to
this organization. We excluded 100 percent wholly owned subsid-
iaries of large U. S. firms (Microsoft, AT&T, Hewlett-Packard,
etc.), as well as the largest Indian firms (TCS, Wipro, Infosys). We
interviewed half of the remaining firms. We picked the firms
randomly (in alphabetical order, skipping over one firm at a time).
We contacted the firms by telephone and arranged an appoint-
ment with the CEO, usually in the following week.
The response rate was excellent. No CEO whom we contacted
refused to meet with us. In Bangalore, where we spent the longest
period of time, all the interviews were arranged. In the other
centers, we had to replace a few firms due to unavailability of the
CEO or another senior person during the time we spent there.
Although nonavailability is in principle not random, we believe
that this was not a big enough problem to significantly affect the
randomness of the sample. The total sample consisted of 125 firms
and 230 contracts.
The interview was conducted with the CEO in most cases, or
with another senior person in the firm. We collected general
information about the firm, and specific information about the last
two projects completed by the firm using a structured question-
naire, available on our web pages. The important limitation is
that we could not collect precise information on the client, data

1016
QUARTERLY J OURNAL OF ECONOMICS
profits, or price charged per man-months by the firm. The project
managers involved in the two projects discussed were often
present as well, or contacted by telephone for classifications. The
interviews lasted from an hour to an hour and a half.
The first part of the questionnaire was straightforward and
was answered by the respondent without any need for verifica-
tion. The project-specific information was either recalled by the
respondent or checked from records. The projects had been
completed very recently, and the information we collected is part
of the data routinely collected by the firms in the course of the
project. They therefore had no difficulty answering these questions.
To facilitate the interview and obtain comparable answers
from firm to firm at each interview, as we were conducting the
interview we drew a timeline representing the different phases of
the project (proposal, specifications, development). We used the
timeline to ask them what the initial estimate was, when it was
revised (after the specifications, at what point in the project), and
what the actual effort was. Several questions can be used to
determine the actual amount of the overruns. First, we asked
them how the actual figure compared with the estimate (making
sure whether they referred to the last estimate or the initial
estimate), and the reasons for these changes. Second, we asked
them separately whether there was any delay due to the client,
any delay due to internal difficulty, and any request for changes
(and when). In the course of asking these questions, we also asked
them whether there was any overrun due to each of these reasons,
and we made sure that the numbers corresponded to what they
had told us before. If they did not, we probed until we reached
consistency (which was usually easily done).
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
REFERENCES
Bajari, Patrick, and Steven Tadelis, ‘‘Incentives versus Transaction Costs: A
Theory of Procurement Contracts,’’ mimeo, Stanford University, 1999.
Crocker, Keith, and Kenneth Reynolds, ‘‘The Efficiency of Incomplete Contracts:
An Empirical Analysis of Air Force Engine Procurement,’’ RAND J ournal of
Economics, XXIV (1993), 127–146.
Diamond, Douglas W., ‘‘Reputation Acquisition in Debt Markets,’’ J ournal of
Political Economy, XCVII (1989), 828–862.
Fama, Eugene F., ‘‘Agency Problems and the Theory of the Firm,’’ J ournal of
Political Economy, LXXXVIII (1980), 288–307.
Heckman, J ames, ‘‘Sample Selection Bias as a Specification Error,’’ Econometrica,
XLVII (1979), 153–161.
Holmstrom, Bengt, ‘‘Moral Hazard in Teams,’’ Bell J ournal of Economics, XIII
(1982), 324–340.

A STUDY OF THE INDIAN SOFTWARE INDUSTRY
1017
Kandori, Michihiro, ‘‘Social Norms and Community Enforcement,’’ Review of
Economics Studies, LIX (1992), 63–80.
McMillan, J ohn, and Christopher Woodruff, ‘‘Interfirm Relationships and Informal
Credit in Vietnam,’’ Quarterly J ournal of Economics, CXIV (1999), 1285–1320.
Pressman, Roger S., Software Engineering, A Practitioner ’s Approach, 4th edition
(New York, NY: McGraw-Hill Companies, 1996).
Tirole, J ean, ‘‘A Theory of Collective Reputations (with Applications to the
Persistence of Corruption and to Firm Quality),’’ Review of Economic Studies,
LXII (1996), 1–22.