Full text of DOI:10.1111/0022-1082.00483

THE JOURNAL OF FINANCE • VOL. LVII, NO. 5 • OCTOBER 2002
Information Production and Capital Allocation:
Decentralized versus Hierarchical Firms
JEREMY C. STEIN*
ABSTRACT
This paper asks how well different organizational structures perform in terms of
generating information about investment projects and allocating capital to these
projects. A decentralized approach—with small, single-manager firms—is most likely
to be attractive when information about projects is “soft” and cannot be credibly
transmitted. In contrast, large hierarchies perform better when information can be
costlessly “hardened” and passed along inside the firm. The model can be used to
think about the consequences of consolidation in the banking industry, particularly
the documented tendency for mergers to lead to declines in small-business lending.
IN THIS PAPER, I TAKE UP the following question: How does organization design
influence the process by which capital is allocated to competing investment
projects? I contrast two basic organizational forms. The first is “decentral-
ization,” in which small, single-manager firms choose between relatively few
projects. The second is “hierarchy,” in which large firms with multiple layers
of management evaluate many projects. The goal is to understand what project-
level characteristics lead either decentralization or hierarchy to be the pre-
ferred design.
This question can be given a concrete motivation. Over the last several
years, there has been enormous consolidation in the banking industry, both
in the United States and worldwide. This consolidation has been accompa-
nied by widely voiced concerns that the resulting larger banks will lend less
to small businesses, which are especially dependent on intermediaries for
financing. And indeed, a number of researchers have documented that, when
two banks merge, the resulting larger entity tends to cut back significantly
on its small-business lending. Moreover, the evidence suggests that the loans
that are cut are, at least on average, positive-NPV.1
Why would a newly enlarged bank ever turn its back on a profitable line
of business in this way? An informal argument is given by Berger, Demsetz,
and Strahan ~1999, pp. 165–166!:
*Stein is with the Department of Economics, Harvard University and NBER. I am grateful
to the National Science Foundation for research support, as well as to Ulf Axelson; Antoine
Faure-Grimaud; Luis Garicano; Rob Gertner; Rick Green; Bengt Holmstrom; Anil Kashyap;
Raghu Rajan; Julio Rotemberg; an anonymous referee; and seminar participants at Harvard
Business School, the NBER, and the University of Chicago for their input.
1
The literature on this topic is surveyed in Berger, Demsetz, and Strahan ~1999!. I discuss
this work in detail in Section III.
1891

1892
The Journal of Finance
The larger institutions created by consolidation may also choose to pro-
vide fewer retail services to small customers because of Williamson ~1967,
1988! type organizational diseconomies. . . . it may be scope inefficient
for one institution to produce outputs which may require implementa-
tion of quite different policies and procedures. These diseconomies may
be most likely to arise in providing services to informationally opaque
small businesses for whom intimate knowledge of the small business, its
owner and its local market gained over time through a relationship with
the financial institution is important . . . these arguments do not sug-
gest that large complex financial institutions created by consolidation
would reduce services to all small customers, rather just to those cus-
tomers who rely on relationships.
On the one hand, this informal argument is quite clear in asserting that
there exist “organizational diseconomies” that somehow prevent big banks
from being the most efficient providers of certain information-intensive ser-
vices, such as relationship-based small business lending. On the other hand,
it is vague as to what the root cause of these diseconomies might be. For
example, the suggestion that big banks simply have trouble engaging in
multiple activities that require different technologies ~“different policies and
procedures”! seems less than compelling. After all, most big banks are in-
volved in a wide range of technologically distinct activities, from check pro-
cessing to credit cards to foreign-exchange trading.
So what is it about small-business lending—as opposed to other banking
activities—that might lead it to be a particularly poor fit for a large banking
firm? In what follows, I argue that the key distinguishing characteristic of
small-business lending is that it relies heavily on information that is “soft”—
that is, information that cannot be directly verified by anyone other than
the agent who produces it. For example, a loan officer who has worked with
a small-company president may come to believe that the president is honest
and hardworking—in other words, the classic candidate for an unsecured
“character loan.” Unfortunately, these attributes cannot be unambiguously
documented in a report that the loan officer can pass on to his superiors.
This situation contrasts sharply with, for example, an application for a home
mortgage loan. Here the decision of whether or not to extend credit is likely
to be made primarily based on “hard,” verifiable information, such as the
income shown on the borrower’s last several tax returns.2
With soft information, the advantage of decentralization is that it strength-
ens the research incentives of line managers. Under full decentralization, a
line manager is also the CEO of his firm and, as such, has the authority to
allocate the firm’s funds as he sees fit. Given that he can count on having
some capital to work with, he knows that his research efforts will not be
wasted, and, hence, his incentives to do research are relatively strong. Said
2
The accounting literature has long drawn a similar distinction between hard and soft in-
formation. See, for example, Ijiri ~1975! and Demski ~1994! for discussions and further references.

Information Production and Capital Allocation
1893
differently, decentralization rewards an agent who develops expertise by
ensuring that he will also have access to some capital that he can use to
lever that expertise.
In contrast, if a line manager works inside a large hierarchy, he faces the
risk that somebody higher up in the organization will decide that invest-
ment opportunities are better elsewhere in the firm and will sharply cut his
capital allocation. In this case, because he does not get a chance to act on the
information that he has produced ~and because he is unable to credibly pass
it on!, the line manager’s research effort goes to waste. Ex ante, this implies
that he does less research in a hierarchical setting. Here the authority to
allocate capital is separated from expertise, which tends to dilute the incen-
tives to become an expert.
Thus, with soft information, the advantage of decentralization relative to
hierarchy is higher-powered research incentives and better capital allocation
within operating units. However, there is also a countervailing cost: Under
decentralization, reallocations across operating units have to be mediated by
the external capital market, while under a hierarchical design, they are im-
plemented by the integrated firm’s CEO. And there can be circumstances in
which a CEO is able to bring higher-quality information to bear on such
across-unit reallocation decisions than the external market does. Indeed, a
central assumption throughout the analysis is that the CEO of an integrated
firm has an advantage over the capital market in learning about the pros-
pects of the units that she oversees. As I argue in detail below, this assump-
tion can be rationalized in a number of ways, perhaps most convincingly by
appealing to a complementarity between the CEO’s strong control rights and
her incentives to produce information.
In sum, when information is soft, decentralization has both advantages
and disadvantages. Balancing the two, decentralization will on net tend to
be a good design under soft information to the extent that line-manager
research—and the accompanying potential for efficient within-unit reallo-
cations—is relatively valuable.
Things work very differently when the information produced by line man-
agers can be hardened and passed on to their superiors. Now, not only does
a hierarchy do better than the external capital market in terms of moving
money across operating units, it can also generate more research on the part
of line managers, and, hence, better within-unit allocations. This is because
with hard information, line managers become advocates for their units; if
they can produce verifiable positive information and pass it on to their su-
periors, they can increase their capital budgets. Here, paradoxically, sepa-
rating authority from expertise actually improves research incentives, as
line managers struggle to produce enough information to convince their bosses
that they should get more of the firm’s resources.
Beyond just saying that soft information favors small firms, the model
also produces several other conclusions. First, suppose that for some other
exogenous reason, it becomes optimal to have a relatively large integrated
firm operating in a setting where information is soft—say, because there are

1894
The Journal of Finance
significant synergies across the firm’s different projects. In such a case, hold-
ing the firm’s size and scope fixed, the softness of information will tend to
imply that a flatter organizational structure, with fewer layers of manage-
ment, is more attractive.
Another implication is that hierarchies tend to be characterized by inef-
ficient levels of bureaucracy. This follows if one extends the model so that
the hardness of information is made endogenous. Suppose that by devoting
effort to documentation, a line manager can harden information that would
otherwise be soft. Because hard information is so privately valuable to line
managers in a hierarchy, they will devote excessive efforts to such documen-
tation. Thus, in this modified setting, the costs of a hierarchy do not take the
form of line managers simply slacking off; rather, they work hard to gener-
ate the wrong kind of information. In particular, there will be too much
report writing and not enough soft-information production.
The ideas in this paper build on several earlier works. I defer a full dis-
cussion of the related literature until Section IV and only note here the most
direct linkages. The capital-allocation model itself is a direct extension of
that in Stein ~1997!, with the new twist being the explicit consideration of
line-manager research incentives. In Stein ~1997!, line managers are treated
as passive robots who do not need to be motivated. Rather, the focus is more
narrowly on the CEO’s incentives to shift resources across the firm’s differ-
ent operating units—an approach that naturally leads to a more favorable
view of large organizations.
In arguing that line managers’ incentives may be blunted when they are in
a hierarchy and do not have ultimate authority, I am following Aghion and
Tirole ~1997!. However, a key distinction is that in the model of this paper,
a hierarchical structure need not weaken line-manager incentives—indeed,
it only does so when information is soft. In contrast, in Aghion and Tirole, it
is a more general proposition that line managers are discouraged when they
do not have full authority. Thus, the models have different empirical impli-
cations, as the example of the banking industry suggests: The Aghion–Tirole
model does not explain why large banks might be particularly unsuited to
small-business lending, as opposed to credit-card or mortgage lending.
The remainder of the paper proceeds as follows. The basic model is devel-
oped in Section I. Section II considers several extensions and variations. In
Section III, I return to the banking industry and review the relevant empir-
ical evidence more fully in light of the theory. Section IV discusses the re-
lated theoretical literature on organizational design. Section V concludes.
I. The Model
A. Basic Structure
The model considers two operating units, i and j, which may be organized
either as two separate stand-alone firms or as subsidiaries under the same
roof of a single integrated firm. Abusing terminology slightly, I will refer to

Information Production and Capital Allocation
1895
these operating units as “divisions” in both the decentralized and hierar-
chical cases. Within each division, there are two potential investment projects.
Each of the projects can be allocated either zero, one, or two units of capital,
and each can be in either a G ~good! or B ~bad! state. The probability of
either state is 102, and the outcomes are independent across all projects,
both within and across divisions.
A project that is in the G state yields a verifiable net output of g~1! . 0 if
it gets one unit of capital and g~2! . 0 if it gets two units. A project that is
in the B state yields a net output of b~1! with one unit of capital and b~2!
with two units. I assume throughout that g~2! Ͻ 2g~1!; that is, there are
decreasing returns in the good state. In addition, g~2! Ͼ g~1! ϩ b~1!, which
means that if one has two units of capital along with a G project and a B
project, it is better to give the G project both units, as opposed to dividing
the capital up equally across the projects.
An innocent normalization that reduces notational clutter, which I adopt
from this point on, is to set b~1! ϭ 0. Finally, I assume that Ϫ1 , b~2! , 0,
as well as that b~2! ϩ g~2! , 0. That is, the net return to investing two units
in a bad project is so negative that it offsets the gains from investing two
units in a good project. As will become clear below, this assumption ensures
that there will be credit constraints in equilibrium, since an uninformed
provider of external finance will never want to give each division four units
of capital to work with. And financing frictions are necessary in this model
if one is to pose questions of internal capital allocation; without such fric-
tions, all projects would get fully funded in all states of the world.3
Each of the two divisions has its own division manager. The division man-
agers are, by virtue of research effort, able to obtain signals about the projects
that they oversee. If manager i makes an effort ei, he has a probability p~ei!
of observing signals on both of his projects, where the function p~ ! is in-
creasing, concave, and takes on values on the interval @0, 1!. I assume that
the division managers have reservation utilities of zero, so that there is
never any issue of satisfying their participation constraints. In the hierar-
chical setting, when the two divisions are integrated into a single firm, the
firm also has a CEO, who may also undertake her own research. The CEO’s
research technology is described in detail below. For the time being, it suf-
fices to say that, because the CEO is overseeing a total of four projects, she
is unable to learn as much about each one individually as are the division
managers.
To model the incentives of the divisions managers, as well as the CEO ~when
there is one!, I assume that any agent k has a utility function of the form:
Uk ϭ ~ yk ϩ Ik! Ϫ gek,
~1!
3
In my model, the firm invests everything it raises from the outside market, due to the
CEO’s empire-building tendencies. By contrast, in papers such as Antle and Eppen ~1985! and
Harris and Raviv ~1996!, there can be internal capital rationing even absent external con-
straints, as the CEO, acting as a principal, seeks to curb the information rents of her subordinates.

1896
The Journal of Finance
where yk is the expected net output of agent k’s division ~or of the whole
firm, if k is the CEO!, Ik is the amount initially invested in the division ~or
in the whole firm, if k is the CEO!, and g is a measure of the degree of effort
aversion. Thus, each agent seeks to maximize the expected gross output
from the assets under his or her control, as given by ~ yk ϩ Ik!, less the cost
of effort. This assumption can be motivated based on: ~a! private benefits of
control that are proportional to gross output, and ~b! nonresponsiveness of
agents to monetary incentives.4 It has the following behavioral implication:
Each agent prefers more capital to less, but conditional on being granted a
certain budget, tries to allocate it efficiently. Or said differently, the agents
in the model are empire builders, but holding the size of their empires fixed,
they prefer them to be profitable.
The goal of the subsections that follow is to establish two principal results.
First, when information is soft, decentralization may for some ~though not
all! parameter values be the more efficient mode of organization. Second,
when information is hard, hierarchy always dominates decentralization. To
proceed, I begin by analyzing the case of decentralization and evaluating the
net surplus created by investment. Next, I turn to the case of a hierarchy
with soft information and again do a similar evaluation, which then allows
me to compare the two regimes. Finally, I take on the case of a hierarchy
with hard information and again compare the net surplus created by invest-
ment in this setting to that under decentralization.
B. Decentralization
B.1. External Financing Constraints Under Decentralization
The analysis of the decentralized case can be broken into two parts. First,
one has to determine how much a decentralized firm is able to raise from
the external capital market. And second, one has to figure out how effi-
ciently this capital will be allocated internally. I assume that investors in
the external capital market cannot directly observe the signals ~if any! re-
ceived by division managers.5 And given that division managers always pre-
fer more capital to less, and are not responsive to monetary incentives, there
is no incentive-compatible way for outside investors to get them to reveal
their private information—division managers will always want to claim that
both of their projects are in the G state, so as to get more financing. Thus,
the best that outside investors can do is to give each division a fixed, un-
contingent allocation of capital. Depending on parameter values, this allo-
4
Nonresponsiveness to monetary incentives is a common—albeit extreme—modeling short-
cut. It can be generated by assuming that agents are infinitely averse to risk in their monetary
income, though not to variations in private benefits. See Aghion and Tirole ~1997!.
5
This assumption is a natural one when division managers’ signals are soft information. It
requires more careful elaboration when division managers’ signals are hard information. This
issue is taken up in detail in Section I.D below.

Information Production and Capital Allocation
1897
cation will be set at either two or three units per division; an allocation of
four units per division can never be optimal given the assumption that b~2! ϩ
g~2! , 0.
I start by simply assuming that each division gets an allocation of two
units from the external market. I then solve for expected output, given di-
vision managers’ endogenously determined incentives in this case. Next, I
repeat the analysis, assuming instead that each division is able to raise
three units of external capital. Then it is a simple matter of comparing net
output across the two cases and checking which leads to higher ex ante
returns for investors.
B.2. Incentives and Output with Two Units of Capital per Division
If a division manager with two units of capital is successful in his research
efforts and observes $G, B% for his two projects, he will behave efficiently ex
post and give both units to the G project. The only question is how much
effort he will put into research ex ante. To answer this question, note that if
a division manager knows the states and allocates the two units based on
this knowledge, expected net output is ~g~2! ϩ g~1!!02. ~Again, recall that
b~1! is normalized to zero.! If, on the other hand, the manager is un-
informed, each project always gets one unit of funding, and expected net
output is simply g~1!. Thus, with a budget of two units, the manager’s utility
gain from being informed under decentralization, denoted by ⌬d2, is
⌬d2 ϭ ~g~2! Ϫ g~1!!02.
~2!
Given equations ~1! and ~2!, the manager’s first-order conditions imply
that, with two units of financing, the level of research effort under decen-
tralization, e2d, satisfies
p'~e2d! ϭ g0⌬d2.
~3!
With two independent divisions each behaving this way, the total expected
net return on four units of capital under decentralization, denoted Y d~4!, is
given by
Y d~4! ϭ p~e2d!~g~2! ϩ g~1!! ϩ ~1 Ϫ p~e2d!!~2g~1!!.
~4!
B.3. Incentives and Output with Three Units of Capital per Division
The analysis of the case where each division has three units of capital is sim-
ilar. If a division manager knows the states and can allocate the three units
conditional on this knowledge, expected net output is ~3g~2! ϩ g~1! ϩ b~2!!04.
Note that in this case, even if the manager is informed, there are some states—
that is, those with two B projects—where it is impossible for outside inves-
tors to avoid the negative b~2! outcome.

1898
The Journal of Finance
If the manager is uninformed, he must randomly allocate two units to one
of the projects, and one to the other, which leads to expected net output
of ~g~2! ϩ g~1! ϩ b~2!!02. Thus, with three units of financing per division, the
utility gain to being informed under decentralization, denoted by ⌬d3, is
given by
⌬d3 ϭ ~g~2! Ϫ g~1! Ϫ b~2!!04.
The level of research effort with three units of financing, e3d, is given by
p'~e3d! ϭ g0⌬d3.
~6!
~5!
With two independent divisions each behaving this way, the total net re-
turn on six units of capital under invested under decentralization, Y d~6!, is
given by
Y d~6! ϭ p~e3d!~3g~2! ϩ g~1! ϩ b~2!!02 ϩ ~1 Ϫ p~e3d!!~g~2! ϩ g~1! ϩ b~2!!. ~7!
A comparison of equations ~2! and ~5! implies that research incentives will
be stronger when a division is allocated three, rather than two units of cap-
ital, if b~2! ϩ g~2! Ϫ g~1! , 0. This condition is always satisfied given the
initial assumptions that b~2! ϩ g~2! , 0 and g~1! . 0. Intuitively, being
informed is more valuable when there are three units of capital to allocate,
because now it helps avoid the very adverse outcome where two units are
put in a bad project.
Of course, it does not follow that allocating three units to each division is
the ex ante optimal policy for outside investors. Indeed, as can be seen by
comparing equations ~4! and ~7!, if b~2! is sufficiently negative, investors
will prefer to only give two units to each division. Although the exact cutoff
value of b~2! for which this is the case depends on the shape of the p~ !
function, it is easy to establish the following sufficient condition ~see the
Appendix for details!
LEMMA 1: If 3~g~2! Ϫ g~1!! ϩ b~2! Ͻ 0, outside investors will always choose to
allocate two units of capital to each division, regardless of the shape of the
p~ ! function.
In much of what follows, I assume that the sufficient condition in Lemma
1 holds, thereby focusing on regions of the parameter space for which equa-
tion ~4! summarizes the output effects of decentralization. This is done solely
for expositional purposes—it cuts down on the number of different cases to
be considered, without changing the important conclusions.

Information Production and Capital Allocation
C. Hierarchy: The Case of Soft Information
C.1. The CEO’s Research Technology
1899
The next regime to be considered is one in which there is soft information
and the firm is organized as an integrated hierarchy, with the two division
managers ceding formal authority to the CEO. To make this case interest-
ing, one needs to assume that the CEO can gather some information on her
own. If not, she can do no better than to always grant each division two
units of funding, thereby reproducing the decentralized outcome.
But allowing the CEO to gather information raises an important question:
Why is it that the CEO is able to become better informed than external
providers of capital, who are assumed to be incapable of learning anything
about the projects? There are two possible answers to this question. First, if
there are multiple outside investors ~e.g., dispersed shareholders!, free-
riding problems will naturally reduce their incentives to acquire informa-
tion. Second, and more subtly, even a single outside investor with the same
research capabilities as the CEO may endogenously choose to do less infor-
mation gathering, to the extent that this outside investor has weaker control
rights than the CEO. That is, there is a complementarity between authority
~either formal or informal! and research incentives.6
As an example, consider the differing incentives of a bank lender versus a
CEO when both contemplate investing effort in learning about a specific
division. The bank can use any information it acquires to guide its sub-
sequent decision of how much to lend to the division. But suppose that in the
course of its investigation, the bank also learns that the division would be
more valuable if its manager were replaced, or if some of its other existing
physical assets were reconfigured. Outside of default, the bank does not
have the authority to impose such outcomes on a reluctant division manager.
In contrast, a CEO overseeing the division can. As a result, the CEO has
more to gain from acquiring information about the division in the first place,
and hence does more research, even if her research technology is no different
from that of the banker.
Nevertheless, even if it is plausible that the CEO learns more about in-
vestment prospects than outside investors, it would be unreasonable to posit
that she can learn as much in total as the two division managers. Instead, I
assume that the CEO only gets coarse information about the aggregate pros-
pects of each division. Specifically, there is a probability q that the CEO’s
research efforts will be successful. Successful research means that, if one or
both divisions are “stars”—in the sense of having both of their projects in
the G state simultaneously—this star status is revealed to the CEO. The
coarseness of the CEO’s research technology is captured in the fact that,
even if her research is successful, she can never differentiate between a
6
This point is modeled by Gertner, Scharfstein, and Stein ~1994!.

1900
The Journal of Finance
division that is “average” ~has one G and one B project! and a division that
is a “dog” ~has two B projects!.
A few points about this formulation deserve comment. First, the CEO’s
research-success probability q is, for the time being, an exogenous param-
eter. In Section II.B, I discuss what happens when q is made an endogenous
function of the CEO’s effort. In either case, I do not allow the CEO to pre-
commit to not doing research ~i.e., to setting q ϭ 0!, even if this precommit-
ment might raise ex ante expected net output.7 Second, the exact way that
I have modeled the coarseness of the CEO’s information is not critical. I
could equivalently assume that successful research allows the CEO only to
identify dog divisions and that she can never distinguish between average
divisions and stars; this leads to the same results. Finally, I am assuming
that the outcome of the CEO’s research effort is independent of the outcome
of the division managers’ efforts. This simplifies the exposition slightly, but
is not necessary. Similar conclusions emerge if, for example, the CEO is
more likely to succeed in learning something conditional on the division man-
agers having also been successful in their research.
C.2. External Financing Constraints in the Hierarchical Case
As before, a complete analysis of the hierarchical case involves solving for
both the optimal amount of funding given to the hierarchy by the external
capital market as well as the incentives which govern how these funds are
allocated internally.8 And things are more complex now, because without
further restrictions on parameters, there are four possible funding levels to
be considered: four, five, six, or seven units. Even if the condition in Lemma 1
is satisfied, so that the two divisions can each raise only two units in the
decentralized case, it is possible that when they are integrated, the hierar-
chy will be able to raise five or more units. This is due to the “ease-the-
credit-constraint” effect identified by Stein ~1997!. Intuitively, because of the
diversification inherent in a larger internal capital market, there is a lower
probability that an extra unit of financing will be invested in a bad project,
leading to the negative b~2! outcome.
Tackling all the scenarios corresponding to the various possible funding
levels is, as before, straightforward. However, such a detailed treatment of
external credit constraints does not serve to advance the main goal here,
which is to show that—in spite of hierarchy’s potential for beneficial across-
division reallocations—with soft information, it is possible for some param-
eter values to have decentralization be the preferred organizational form. To
illustrate this point in the most transparent way, I impose another restric-
7
This no-precommitment assumption is discussed further in Section II.A. Cremer ~1995! is
another paper in which a supervisor’s inability to commit to staying uninformed can weaken
the incentives of the agent she supervises.
8
The feature that no information can be credibly revealed to the external capital market still
remains. Since the CEO also derives private benefits that increase with total investment, she
will have the same incentive to misrepresent investment prospects as the division managers.

Information Production and Capital Allocation
1901
tion on the parameters that is sufficient to ensure that a hierarchy will
never be able to raise more than four units of funding. This restriction puts
aside the ease-the-credit-constraint effect, thereby isolating the trade-off
between giving a CEO the ability to make transfers across divisions, versus
diluting the research incentives of the division managers. In the Appendix,
I prove the following lemma.
LEMMA 2: If 25~g~2! Ϫ g~1!! ϩ b~2! Ͻ 0, outside investors will only give four
units of capital to a hierarchy, regardless of how efficiently the hierarchy
allocates this capital internally.9
Again, the point is not that the condition required in the lemma—
effectively, that the negative b~2! outcome is very bad—should be thought of
as representing the most likely case. Rather, it enables one to focus on the
simplest example of the costs and benefits of hierarchy, without worrying
that this example is inconsistent with optimization on the part of investors.
C.3. Incentives and Output When the Hierarchy
Has Four Units of Capital
Given that it is able to raise four units of external finance, capital allo-
cation in a hierarchy works as follows. When the CEO’s research is unsuc-
cessful ~which happens with probability ~1 Ϫ q!!, the best she can do is to
just give each division manager two units of funding and we are back to the
corresponding decentralized outcome.10 When the CEO’s research is success-
ful ~which happens with probability q!, she may choose to deviate from equal
funding and give one division more than the other. This will only happen if
one division is identified as a star, and the other is not. In such a “lone-star”
scenario, the CEO has three options: ~a! still give each division two units,
~b! give the star division three units and the other division one unit, or
~c! give the star division all four units. It is easy to show that the CEO will
choose the most extreme tilting of the capital budget—giving all four units
of funding to the star division—if the following sufficient condition is met:
g~2!02 Ͼ 5g~1!06.
~8!
9
In proving the lemma, I implicitly rule out schemes where the CEO raises four units of
financing, irrevocably sinks it all in division i, and then later returns to the capital market
again, asking for two more units for the as-yet-unfunded division j. One way to rationalize this
restriction is to assume that outsiders cannot verify whether investment expenditures “belong”
to one division of an integrated firm versus another; such an assumption is a common one in
the literature on internal capital allocation. See Scharfstein and Stein ~2000! for a discussion.
10
I do not consider mechanisms that the CEO might use to induce division managers to
reveal their signals. One could imagine that if the CEO catches a manager lying about his
signal, he would be punished with a reduced capital allocation, as in Harris and Raviv ~1996!.
In my setup, such a scheme suffers from two distinct commitment problems. First, punishment
involves ex post inefficient allocations. And second, if in equilibrium the division managers do
truthfully reveal their signals, the CEO will no longer have any incentive to do her own research.

1902
The Journal of Finance
This condition simply requires that decreasing returns to scale are not too
pronounced in the good state. The analysis that follows is most intuitive
when this star-gets-everything condition is satisfied, so I will begin by assum-
ing that it is. Later, I will come back to the case where returns decrease
more sharply with scale, so that the CEO gives just three, rather than four
units of funding to a division that is identified as a lone star.
If division managers’ ex ante research incentives in a hierarchy were the
same as under decentralization, it would follow that hierarchy is the strictly
dominant organizational form. This is because ex post, hierarchy allows for
a form of selective intervention. When the CEO knows nothing about divi-
sional prospects, she does not interfere, and the outcome is the same as with
decentralization. When the CEO does know something, her ability to shift
funds towards a star division leads to an improved across-division allocation.
The problem, however, is that division managers’ ex ante research incen-
tives are weaker in a hierarchy when information is soft. To see why, sup-
pose that the CEO’s research has been successful and that she has identified
division j as a star. Division i, meanwhile, has one G and one B project. In
a hierarchy, division j gets all four units of funding, and division i gets noth-
ing. Hence, any information that manager i has acquired is not put to use.
In contrast, if the divisions were decentralized, and manager i had two units
of funding to work with, he would find his information valuable—it would
lead him to shift both units to his single G project. Thus, the downside to a
hierarchy is that because the CEO sometimes takes away manager i’s cap-
ital budget, the marginal return to his research effort is reduced, and he
produces less information.11
It is important to recognize that the negative incentive effects of hierarchy
arise not simply because the CEO sometimes has her own independent in-
formation about divisional investment opportunities. It is also crucial to the
argument that the CEO have the authority to take away all funding from
division i—even though i would be able to raise two units if it were a stand-
alone entity—when her research indicates that division j is a star.12 As em-
phasized by Stein ~1997!, this authority distinguishes the CEO from an equally
well-informed outside provider of finance, such as a banker.
11
This discouragement effect is not offset by the fact that the manager also sometimes gets
two extra units in a hierarchy. When he receives four units, the only thing he can do is invest
two in each project, and information is again not of any value. Moreover, the feature that a
stable capital budget creates stronger research incentives than a variable one is actually more
general than the simple setup here suggests. Suppose instead that investment is continuous,
that an amount I yields f ~I! in state G, 0 in state B, and that a manager has one project of
each type. If he is uninformed and has a budget of K, he will invest K02 in each project, yield-
ing f ~K02!. If he is informed, he will get f ~K!. Hence, the value of being informed is V~K! ϭ
f ~K! Ϫ f ~K02!. A stable capital budget will lead to more research if V~K! is concave, or if
4f ''~K! Ϫ f ''~K02! Ͻ 0. This mild condition is satisfied, for example, for any f ~I! of the power
form f ~I! ϭ I a0a, for 0 , a , 1.
12
Indeed, the discouragement effect occurs precisely because an “average” division ~i.e., one
with one G and one B project! can lose all its funding based not on its own prospects, but rather
because the prospects of its counterpart division within the firm are so strong.

Information Production and Capital Allocation
1903
Assume that the timing of the game is such that a division manager must
choose his level of research effort before he knows whether the CEO’s re-
search has been successful. Denote by ⌬h4s the gain in expected utility that an
individual division manager obtains when, in a hierarchy with soft informa-
tion and four units of capital, his own research efforts succeed. It is straight-
forward to show that
⌬
hs ϭ ~1 Ϫ q!~g~2! Ϫ g~1!!02 ϩ 3q~g~2! Ϫ g~1!!08 ϭ ~1 Ϫ q!⌬d2 ϩ 3q⌬d204. ~9!
4
The corresponding level of research effort, e4hs, satisfies
p'~e4hs! ϭ g0⌬h4s.
~10!
Since ⌬h4s Ͻ ⌬d2, it follows that e4hs Ͻ e2d. By working through all the possible
outcomes, it can then be established that expected net output in a hierarchy
with soft information and four units of capital is given by
Y hs~4! ϭ ~1 Ϫ q!$ p~e4hs!~g~2! ϩ g~1!! ϩ ~1 Ϫ p~e4hs!!~2g~1!!%
~11!
ϩ q$ p~e4hs!~6g~2! ϩ g~1!!04 ϩ ~1 Ϫ p~e4hs!!~3g~2! ϩ 4g~1!!04%.
By comparing equations ~11! and ~4!, one can evaluate the relative effi-
ciency of decentralization versus hierarchy, conditional on there being a to-
tal external capital constraint of four units in either case. The results to this
point can be summarized in the following proposition.
PROPOSITION 1: Assume that the condition in Lemma 2 holds, so that the ex-
ternal credit constraint is four units. Assume further that inequality (8) holds,
so that in a hierarchy, a lone-star division gets all four units of funding.
Then decentralization always leads to more research effort than hierarchy:
e4d Ͼ e4hs. In addition, it is possible ~though not necessary! that decentraliza-
tion leads to higher expected output; that is, that Y d~4! Ͼ Y hs~4!.
To see why decentralization can generate higher expected output, consider
a simple limiting case where q ϭ 1, p~e4d! ϭ 1, and p~e4hs! ϭ 0. ~The latter two
conditions can always be generated by picking the proper form for the p~ !
function.! In this case, equation ~4! simplifies to Y d~4! ϭ g~2! ϩ g~1!, while
equation ~11! simplifies to Y hs~4! ϭ ~3g~2! ϩ 4g~1!!04, implying that Y d~4! Ͼ
Y hs~4!. In this example, when the firm is organized as a hierarchy, only the
CEO does any research, and division managers are totally discouraged. Con-
versely, under decentralization, the division managers are highly motivated
and become perfectly informed. Given that the two division managers taken
together are able to gather more accurate information than the CEO, this
latter effect is more than enough to outweigh any improved across-division
allocation that can be obtained in a hierarchy. As a result, decentralization
is the better mode of organization.

1904
The Journal of Finance
Of course, this example relies on p~e4d! and p~e4hs! being relatively far
apart—that is, on division-manager effort being both important and respon-
sive to incentives. If p~e4d! and p~e4hs! are sufficiently close to one another, it
is easy to see that hierarchy becomes more efficient than decentralization:
Y hs~4! Ͼ Y d~4!.
Although I have derived the results in Proposition 1 under the assumption
that the star-gets-everything condition in ~8! holds, they are, in fact, more
general. Even when returns to scale are more sharply decreasing in the good
state—so that a lone-star division gets three, rather than four units of
funding—the basic intuition can carry over. That is, when manager i gener-
ates information and gets allocated only one unit of financing ~because man-
ager j has been deemed a star by the CEO!, manager i’s information is not
worthless, but it may still be less valuable than if he had been allocated two
units of financing. Thus, the threat of losing some—if not all—of their fund-
ing can continue to exert a negative effect on division managers’ research
incentives.
To make this idea precise, in the Appendix I prove the following proposition.
PROPOSITION 2: Assume that the condition in Lemma 2 holds, so that the ex-
ternal credit constraint is four units. Assume further that g~2!02 Ͼ 3g~1!04.
In this case, the CEO in a hierarchy will still tilt the capital budget toward
a lone-star division, but the tilt may be less extreme, with the lone star re-
ceiving three, rather than four units of funding. Nevertheless, decentraliza-
tion continues to lead to more research effort than hierarchy: e4d Ͼ e4hs. In
addition, it is possible ~though not necessary! that decentralization leads to
higher expected output than hierarchy, that is, that Y d~4! Ͼ Y hs~4!.
D. Hierarchy: The Case of Hard Information
The next task is to show that the relative merits of a hierarchy increase
when the information generated by division managers can be hardened and
passed along to the CEO. I begin by assuming that the sufficient condition
in Lemma 2 continues to hold. This allows for a simple comparison of hier-
archy relative to decentralization in a situation where the total amount of
available external finance is four units in either case. But once I have es-
tablished that a hierarchy with hard information dominates decentralization
for these parameter values, it is easy to extend the argument to show that it
must dominate for all parameter values—that is, regardless of how many
units of financing can be raised under either decentralization or hierarchy.
To introduce hard information, I assume that the division managers have
the same research technology as before—that is, if manager i makes an
effort ei, he has a probability p~ei! of observing the signals on both of his
projects. Now, however, the CEO does no separate research of her own. In-
stead, if a division manager learns something, there is a chance that he may
be able to credibly communicate it to the CEO. Specifically, conditional on a
division manager’s research being successful and yielding information about

Information Production and Capital Allocation
1905
his two projects, there is a probability z that this information is verifiable
and can be shown directly to the CEO. With probability ~1 Ϫ z!, the infor-
mation is nonverifiable and can be used by the division manager, but not
credibly transmitted to the CEO.
A critical assumption is that while hard information can be credibly trans-
mitted from division managers to the CEO, it cannot be revealed to outside
investors. Otherwise, the hierarchy would no longer face a fixed external
credit constraint. What is the economic interpretation of such an assump-
tion? Think of division managers as providing the CEO with a variety of raw,
albeit well-documented data about a project.13 For example, if the project
involves drilling for oil in a new location, the raw data might be a set of
geological studies. These studies do not literally say what the project’s dollar
payoff will be. Instead, they must be combined with the CEO’s specific ex-
pertise ~e.g., her knowledge of geology, her assessment of the costs of doing
the drilling and extraction! to generate a final judgment about dollar value.
As before, the complementarity of the CEO’s authority and research in-
centives imply that she will in equilibrium be better able to assess such raw
data than would an outside banker. Continuing with the example, because
the CEO has the authority to decide how to deploy the firm’s drilling equip-
ment, she will have more incentive to become informed about the costs of
drilling in different locales, which in turn makes her better able to assess
the value implications of the geological studies. The CEO’s final judgment—
what she concludes after studying the raw data—is itself soft information
that cannot be credibly communicated to outside investors.
Under this interpretation, the parameter z can be thought of as reflecting
either the ability or effort of the CEO—a higher value of z means that the
CEO can better evaluate the raw data produced by the division managers.
Moreover, the idea that the CEO needs to spend time evaluating raw data
helps to motivate the assumption that, when information is hard, she no
longer does her own separate research into divisional prospects: She is too
busy trying to process her subordinates’ output.14 As will be made clear be-
low, the consequence of this assumption is that when information is hard,
CEO and division-manager effort become strategic complements, as opposed
to substitutes.
Another crucial assumption is that the CEO does not observe whether
division-manager research has been successful. Consequently, when a divi-
sion manager has hard information, he can choose whether or not to report it
13
Internal accounting or auditing mechanisms may have a role to play in making such raw
data sufficiently well documented as to be credible.
14
Clearly, this aspect of the model could be more fully endogenized. For example, the CEO
could face a trade-off between time spent gathering her own independent information and time
spent interpreting division managers’ reports, and one could establish the conditions under
which she tended to find the latter optimal. This line of thinking expands on Aghion and Tirole’s
~1997! idea of keeping the CEO too busy to do her own separate project evaluation. Here, the
CEO is kept busy not because she is overloaded with make-work, but rather because she chooses
to engage in another more productive activity.

1906
The Journal of Finance
to the CEO.15 In particular, a division manager who gets hard information
that his projects are $B, B% may opt to simply keep quiet. Importantly, this
will not lead to an “unraveling” situation where the CEO can infer that the
state must be $B, B% simply because the division manager is quiet. This is
because the division manager may also be quiet as a result of not having
obtained any hard information in the first place.
The option to keep quiet makes research very attractive to the division
manager. To see why, assume that the CEO has a fixed conjecture eic about
the level of effort that manager i exerts, and a corresponding conjecture pc ϭ
p~eic! about the probability that his research is successful. Now suppose that
the division manager’s reporting strategy is to reveal his hard information
to the CEO when it is either $G, G% or $G, B%, but to keep quiet when it is
$B, B%. ~As will be seen below, this strategy can constitute equilibrium be-
havior.! Bayes’ rule implies that, given his conjecture of pc, the CEO will
interpret quiet as follows:
prob~$B, B%0quiet! ϭ 10~4 Ϫ 3zpc !
~12!
~13!
~14!
prob~$G, G%0quiet! ϭ ~1 Ϫ zpc !0~4 Ϫ 3zpc !
prob~$G, B%0quiet! ϭ 2~1 Ϫ zpc !0~4 Ϫ 3zpc !.
Taking the CEO’s conjecture of pc as fixed, a division manager reasons as
follows. If he does no research, he will certainly be quiet, and the CEO will
update on him according to equations ~12!–~14!. However, if he does devote
some effort to research, he gains pure option value. If the research produces
hard information that his division is $G, G%, he can speak up and thereby
impress the CEO. If the research produces hard information that his divi-
sion is $B, B%, he just keeps quiet and is no worse off than if he had done no
research.
The bottom line is that, from the perspective of the division manager,
there is now an added benefit to doing research: If the information he gen-
erates is hard and positive, it can help increase his capital budget and hence
his private benefits. This contrasts with both of the previous scenarios—
decentralization and hierarchy with soft information—where division man-
agers’ research efforts had no impact on the capital they were allocated.
In the Appendix, I provide a detailed characterization of the case when
information can be hardened. The key results can be summarized as follows.
PROPOSITION 3: Assume that the condition in Lemma 2 holds, so that the ex-
ternal credit constraint is four units. Assume further that g~2!02 Ͼ 3g~1!04.
Then in a hierarchy with hard information, there is an equilibrium with the
following properties:
15
A similar assumption has been made in accounting research that addresses the topic of
discretionary disclosure to the capital market. See, for example, Verrecchia ~1983! and Dye
~1985!.

Information Production and Capital Allocation
1907
(a) Division manager reporting strategies: Division managers reveal their
hard information if it is either $G, G% or $G, B% and keep quiet if it is
$B, B%.
(b) CEO capital-allocation policy: If the CEO is facing one division that
reveals itself to be $G, G% and one that is quiet, the $G, G% division gets
at least three (and possibly four) units of capital. In all other cases, the
CEO allocates each division two units of capital.
(c) Division manager benefit from being informed: Denote by ⌬h4h the util-
ity benefit to a division manager from being informed in a hierarchy
with hard information. This benefit exceeds that under decentraliza-
tion: ⌬h4h Ͼ ⌬d4 ϩ z04. Consequently, division-manager research effort is
greater than under decentralization: e4hh Ͼ e4d.
(d) Output: Expected net output is greater than under decentralization:
Y hh~4! Ͼ Y d~4!.
The intuition behind parts ~a! and ~b! of the proposition has already been
discussed.16 Part ~c! gives a sense for how much stronger division managers’
research incentives are when information can be hardened. The utility gain
to being informed, ⌬h4h, is now not just larger than that under decentraliza-
tion ⌬d4; it exceeds it by at least z04. This can be a substantial difference,
since ⌬h4h is denominated in units of net returns, while z04 is in units of gross
capital. The difference arises because in a hierarchy with hard information,
research effort can actually influence a division manager’s gross capital bud-
get. In contrast, under decentralization ~or in a hierarchy with soft informa-
tion! research effort only enables a division manager to get a higher return
from a given capital budget. To the extent that the division manager’s utility
is tied to gross output, the former effect is naturally much stronger than the
latter. Part ~d! of the proposition follows from part ~c!. Now a hierarchy does
better than decentralization on both dimensions of importance. Not only does
it generate more information at the division-manager level, and thereby lead
to better within-division allocations, but it also allows for reallocations across
divisions when the CEO learns that such reallocations are value-increasing.
Although Proposition 3 deals with the case where the condition in Lemma
2 holds, so that the hierarchy faces an external credit constraint of four
units, the result can easily be generalized. In the appendix, I prove the fol-
lowing proposition.
PROPOSITION 4: Consider the more general case where the only restriction on
b~2! is that g~2! ϩ b~2! Ͻ 0, while keeping the assumption that g~2!02 Ͼ
3g~1!04. Under decentralization, each division may now be able to raise ei-
ther two or three units of capital, and, under hierarchy, the aggregate credit
16
There are other equilibria that differ insignificantly from that in Proposition 3. When one
division reveals $G, G% and the other reveals $G, B%, the CEO is actually indifferent between
allocating two units to each, or three to the former and one to the latter. I focus on the two-two
allocation in the proposition, which implies that a manager with a $G, B% signal strictly prefers
to reveal it. However, parts ~c! and ~d! of the proposition apply in either case.

1908
The Journal of Finance
constraint may be anywhere from four to seven units. Still, if information is
hard, expected net output is always strictly greater in a hierarchy than under
decentralization.
One might argue that the results in Propositions 3 and 4 in favor of hier-
archy relative to decentralization are now “too strong”—once information
can be hardened, hierarchy dominates decentralization for all parameter val-
ues, which seems patently unrealistic. Certainly, the model omits a number
of other factors that might tip the balance back towards decentralization.
For example, other than research effort, the model assumes that division
managers do not need to take any other actions. If one were to introduce
another dimension of noncontractible firm-specific investment, this might
~following the logic of Grossman and Hart ~1986!, Hart and Moore ~1990!,
and Hart ~1995!! be expected to make hierarchy less attractive. But it is
important to emphasize that I am less interested in making absolute state-
ments about the virtues of hierarchy compared to decentralization, and more
interested in making comparative-statics statements about the circum-
stances under which hierarchy is likely to be attractive. And the key result
in this regard—that hierarchy looks better when information is hard, as
opposed to soft—seems like it should be robust to the inclusion of various
other factors into the model.
II. Further Issues
A. Does Decentralization Require Dis-Integration?
I have been treating the concept of decentralization as synonymous with
formal dis-integration, that is, with the two divisions being operated as in-
dependent stand-alone firms. The key assumption has been that it is impos-
sible for a single firm to replicate the decentralized outcome, because the
CEO cannot commit to either remaining uninformed or to not using her
information to reallocate capital across divisions. Given that the model has
no operating synergies across the two divisions, not much hangs on the dis-
tinction between decentralization and dis-integration. But what if there are
significant synergies? Then one must ask if it is possible to capture the
benefits of decentralization—that is, to commit to division managers that
they will always receive two units of funding—without resorting to a costly
breakup.
On the one hand, it is hard to argue that a CEO can definitively alienate
her right to get involved in capital-allocation decisions. Indeed, several au-
thors ~Williamson ~1975!, Donaldson ~1984!, Stein ~1997!, and Scharfstein
and Stein ~2000!! have argued that the CEO’s authority to move capital
across divisions is the single most defining characteristic of an integrated
firm.17 On the other hand, as Aghion and Tirole ~1997! point out, there are
17
Baker, Gibbons, and Murphy ~1999! go a step further, claiming that formal authority over
all decisions necessarily resides at the top of an organization.

Information Production and Capital Allocation
1909
other devices short of a breakup that can reduce—albeit not fully eliminate—
the CEO’s ex post incentive to meddle in the capital-allocation process; they
suggest keeping the CEO very busy, so she cannot do much research on her
own. In a banking context, a relevant example of such a device might be the
creation of a multibank holding company. By making each of the divisions a
legally distinct entity, such a structure could create some impediments to
moving capital freely among them.18
This logic leads to the following qualitative conclusions. Suppose that based
purely on division-manager research incentives, an analysis such as that in
Section I.C cuts in favor of decentralization. Suppose further that the other,
non-capital-allocation-related synergies from integration are given by X. When
X is relatively small, it will make sense to decentralize “to the max” by
breaking up the firm. However, when X is large, it may be better to keep the
firm integrated, but at the same time to do as much as possible with various
internal devices ~e.g., the multibank holding-company structure! to mimic—
even if one cannot fully replicate—the decentralized outcome in terms of
division-manager incentives.
B. CEO Research Incentives: Why Soft Information Favors
Flatter Organizations
In modeling a hierarchy with soft information, I have thus far taken as
exogenous q, the probability that the CEO’s research will be successful. Now
I ask what happens in the soft-information case when the CEO’s incentives
are also taken into account—that is, when q ϭ q~eCEO !, where eCEO is the
research effort exerted by the CEO, and where q~ !, like p~ !, is an increas-
ing concave function. For simplicity, I stick to the case where the external
credit constraint is four units—that is, where the sufficient condition in
Lemma 2 is satisfied—and where the star-gets-everything inequality ~8! also
holds.
The outcome~s! in this setting are the Nash equilibria of the game where
the CEO and the division managers choose their research efforts simulta-
neously. Denote by ⌬CEO the utility gain to the CEO when, in a hierarchy
with soft information, her research effort is successful. This quantity is eas-
ily calculated by evaluating Y hs~4! in equation ~11! for both q ϭ 1 and q ϭ 0,
and taking the difference:
CEO
⌬
ϭ p~e4hs!~2g~2! Ϫ 3g~1!!04 ϩ ~1 Ϫ p~e4hs!!~3g~2! Ϫ 4g~1!!04.
~15!
An immediate consequence of equation ~15! is that the CEO’s research is
more valuable when division managers do less research—that is, when
p~e4hs! is low. Specifically, differentiation of ~15! yields
d⌬CEO0dp~e4hs! ϭ ~g~1! Ϫ g~2!!04 Ͻ 0.
~16!
18
I am grateful to Raghu Rajan for suggesting this example.

1910
The Journal of Finance
It then follows from the CEO’s first-order conditions that deCEO0de4hs , 0
also: The CEO’s research effort is reduced when the division managers are
working hard.19 The intuition behind this result is straightforward. Recall
that the CEO, even when her research is successful, has at best coarse in-
formation. Thus, while her tilting of the capital budget adds value on aver-
age, it does have a cost in some states of the world. In particular, when faced
with one division that is a star ~i.e., that is $G, G%! and one that is not, the
CEO gives the star all four units of funding. This full tilting of the capital
budget toward the star is optimal if it turns out that the nonstar division is
$B, B%. However, the tilt is too extreme if the nonstar division turns out to be
$G, B%—ideally, it would be better to leave a $G, B% division with at least one
unit of funding.
In other words, because of her coarse information, an activist CEO some-
times takes away too much from a $G, B% division. And what is the cost of
taking capital away from a $G, B% division? It depends on how profitable this
division can be expected to be. If its division manager is informed, the $G, B%
division generates a greater expected return from a given allocation of cap-
ital, as this capital is always steered to the better project within the division.
Thus, the opportunity cost of the CEO’s activism is greater when she runs
the risk of taking resources away from informed division managers.
Note the symmetry that is at work here: the CEO’s research incentives are
blunted by the possibility that the division managers will become informed,
much as the division managers’ incentives are blunted by the possibility that
the CEO will become informed. That is, the more senior agent in the orga-
nization can be discouraged by the hard work of her subordinates, as well as
vice versa. This symmetry has a couple of consequences. First, since the
CEO’s effort eCEO is a decreasing function of division managers’ efforts e4hs
,
and conversely, this game—without further restrictions on the p~ ! and q~ !
functions—admits the possibility of multiple Nash equilibria.20 For example,
one might have either a “control-freak” equilibrium where the CEO is highly
informed and intervenes often, and where the division managers are very
discouraged, or a “laissez-faire” outcome where the reverse occurs. Depend-
ing on parameter values, one equilibrium will be ex ante more efficient than
the other, and there is no guarantee that a firm will end up in the better
one. Thus, for example, the negative consequences of the hierarchical form
of organization may be more pronounced for a firm that has gotten stuck in
a control-freak culture, as opposed to one that has somehow managed to
maintain a laissez-faire environment.
A further implication of the model with endogenous CEO effort is that it
suggests another organizational form that may be optimal in some circum-
stances. In the same way that decentralization can be valuable as a precom-
19
The model of Aghion and Tirole ~1997! produces a similar result.
In other words, Nash equilibrium involves the intersection of two downwards-sloping curves
20
in ~eCEO, e4hs! space. Without any further restrictions on functional forms, it is possible that
these two curves may cross more than once.

Information Production and Capital Allocation
1911
mitment to get the CEO out of the picture—and thereby increase division-
manager incentives—it might sometimes make sense to remove the division
managers, so as to increase CEO incentives. It is easy enough to construct
numerical examples that have this feature. The key is to make CEO effort
both valuable and highly elastic, while making division-manager effort less
so. In such cases, we are left with just the CEO overseeing all four invest-
ment projects, which can be interpreted as an integrated ~i.e., large! firm
with a flat management structure.
Thus, when we compare hard versus soft information in terms of their
implications for organizational form, we now have a new conclusion. Not
only does soft information tend to favor decentralization ~i.e., smaller firms
with fewer projects!, but holding fixed a firm’s size and scope, soft informa-
tion also favors a flatter, more streamlined management structure. This is
because the efforts of higher-ups and subordinates are strategic substitutes
when information is soft, but strategic complements when information is
hard.
C. Hardness of Information Is Endogenous:
Excess Bureaucracy in Hierarchies
I have been assuming throughout that the hardness of information is
exogenous—either a division manager’s information can or cannot be cred-
ibly transmitted to the CEO, but there is nothing that the division manager
can do to influence this. An alternative approach is to posit that the degree
of hardness is endogenous, and, that by expending additional effort on doc-
umentation, a division manager can harden information that would other-
wise be soft.21
To capture this idea formally, I return to the version of the model in Sec-
tion I.D, where the CEO does no research of her own, and make one modi-
fication. Now each division manager can choose between one of two research
technologies: He can either put one unit of effort into acquiring soft infor-
mation, or he can put one unit of effort into acquiring hard information. If
he opts for soft information, there is a probability ps that his research is
successful. If instead he chooses to go after hard information, there is a
probability ph that his research is successful, and conditional on success, a
probability z that the information actually turns out to be verifiable. I as-
sume that ph ϭ bps and b , 1. Thus, effort devoted to acquiring hard in-
formation is less productive. A high value of b means that information about
the project in question is by its nature relatively easy to document, so that
not too much of a price is paid to make it hard; one can think of the previous
21
An example would be an academic department chair, who, having already decided that he
wants to make a tenure offer to somebody, undertakes the process of writing for outside letters.
The letters may provide no new information to the department chair, but they can help him to
credibly sell the case to a less well-informed dean.

1912
The Journal of Finance
cases of absolutely “hard” and “soft” information in Section I as correspond-
ing to the polar extremes where b ϭ 1 and b ϭ 0, respectively.
It is easy to see that even when b is much less than one, the only equi-
librium may be one in which both division managers opt to go after hard
information. If both managers go after soft information, the CEO never learns
anything, and we are effectively in a decentralized outcome, with each divi-
sion manager’s effort yielding him a utility gain of ps⌬d4 ~assuming the ex-
ternal credit constraint is four units!. In contrast, if one division manager
deviates and goes after hard information, we know from part ~c! of Propo-
sition 3 that this manager’s utility gain will exceed ph~⌬d4 ϩ z04!. So even
when ph is much smaller than ps, hard information can yield higher private
benefits. This is just an application of the logic developed above: Hard in-
formation is more attractive to division managers, because it can help them
get larger capital budgets.
Moreover, while hard information is, all else equal, more valuable to the
firm as a whole—it enables the CEO to make value-enhancing reallocations—
division managers’ preference for it is far too strong. That is, there are low
values of b for which the firm would be better off if the division managers
pursued soft information, but division managers’ private incentives are such
that they pursue hard information instead. This is because of an externality.
Manager i’s utility goes up by one when he generates hard information that
lands his division one additional unit of capital. However, manager j loses an
equal amount, and the gain to the firm as a whole is only proportional to the
improved net return on this one unit.
This logic implies that, when project information is innately hard to doc-
ument ~i.e., when b is small but nonzero!, the costs of a hierarchical form of
organization may manifest themselves not just as division managers slack-
ing off and doing no research, as in Section I.C. Rather, one may observe
division managers working extremely hard at creating the wrong kind of
information. In other words, a hierarchy may be characterized by a great
deal of bureaucracy, in the sense of division managers generating lots of
reports that are very well documented, but ultimately not terribly informa-
tive. And conversely, if a hierarchical firm is broken up, the excessive bu-
reaucracy vanishes, and division managers instead produce only soft
information.
III. Empirical Implications: A Closer Look at Banks’
Small-Business Lending Practices
The most basic implication of the theory developed above is that large,
hierarchical firms are at a comparative disadvantage when information about
individual investment projects is innately soft. Moreover, if one takes the
model seriously, the hardness or softness that is most relevant for organi-
zational form has to do with information about those “small” projects that
are overseen by line managers. In other words, what matters is the nature

Information Production and Capital Allocation
1913
of information that, in a hierarchical setting, would be produced far from the
ultimate decisionmaker, the CEO.
While these ideas would seem to have significant empirical content, there
are challenges in mapping the theory into a set of precise, differentiating
predictions that can be readily tested. For example, one interesting impli-
cation of the theory is that small firms might be better able than large ones
to engage in certain types of new product development, since the prospects
of many new products must often be assessed based on soft information.
However, any test of this proposition would have to control for a variety of
other mechanisms that could lead to a similar outcome—for example, the
well-known “replacement effect,” whereby large firms are discouraged from
innovating for fear of cannibalizing their existing product lines.22 Imple-
menting such a control would most likely require a careful case-by-case analy-
sis of the new products in question.
These sorts of complications underscore why it can be particularly infor-
mative to look at small-business lending by banks. Here we have a well-
defined “industry” where: ~a! it is easy to identify the primary “projects” that
line managers must choose among—namely, individual loan applications; and,
moreover, ~b! it seems quite plausible that information about these partic-
ular projects is likely to be innately soft.
The following findings emerge from the empirical literature on small-
business lending. First, small banks invest a much greater share of their
assets in small-business loans than do large banks ~Nakamura ~1994!, Berger,
Kashyap, and Scalise ~1995!, Berger and Udell ~1996!, Peek and Rosengren
~1996!, and Strahan and Weston ~1996!!.23 Perhaps more strikingly, when
large banks acquire small banks, the small-business lending of the combined
organization tends to fall sharply ~Peek and Rosengren ~1998!, Berger et al.
~1998!, Sapienza ~2002!!. Moreover, it appears that the loans that are cut
as a result of consolidation are not cut simply because they are negative-
NPV—that is, because the acquiring bank is cleaning out the bad loans of
the target. Two pieces of evidence support this view. Berger et al. ~1998!
establish that many of the loans that are cut in the process of consolidation
are picked up by other banks in the same local market. And Sapienza finds
that there is no relationship between a borrowing firm’s credit quality and
the likelihood that it will have its lending relationship severed following a
merger.
Although these patterns are broadly consistent with the theory developed
above, they do not pinpoint the exact mechanism at work. That is, they do
not explain why large banks might be disadvantaged at small-business lend-
ing. However, other findings are beginning to emerge that speak more di-
rectly to the central idea of this paper, namely that large organizations are
22
See Tirole ~1988, Chapter 10! for a discussion of the replacement effect.
Relatedly, Brickley, Linck, and Smith ~2000! document that large banks are the dominant
23
players in densely populated metropolitan areas ~where presumably borrowers are more likely
to be big firms!, while small banks have a greater role in suburban and rural regions.

1914
The Journal of Finance
not well suited to handling soft information. Three types of studies are es-
pecially worth noting.
First, it appears that it is not just bank size per se that discourages small-
business lending, but rather organizational complexity. For example, De-
Young, Goldberg, and White ~1997! show that controlling for a bank’s size
and age, its proclivity for making small-business loans is also negatively
related to the number of branches it has, as well as to its being part of a
multibank holding company.24 Keeton ~1995! finds similar results, with a
particularly negative effect on small-business lending for banks that are
owned by out-of-state holding companies.
Second, there is some evidence that in making small-business loans, large
banks tend to shy away from those “difficult credits” where soft information
is likely to be most important in assessing whether or not the loan is positive-
NPV. Berger and Udell ~1996! find that large banks charge about 100 basis
points less on small-business loans than do small banks and require collat-
eral about 25 percent less often. One interpretation is that large banks only
lend to those small customers whose financial position is so strong that
detailed further investigation is not needed.
Finally, Cole, Goldberg, and White ~1997! use a new survey of small-
business finance to look at differences in the loan approval process across
large and small banks. They show that large banks ~over $1 billion in assets!
tend to base loan approvals primarily on standard criteria obtained from
financial statements. In contrast, “small banks deviate from these criteria
more and appear to rely on their impression of the character of the borrower
to a larger extent” ~p. 1!.25 This evidence fits very nicely with the spirit of
the model developed above.
While much of the foregoing discussion has implicitly treated the softness
of information in small-bank lending as an exogenously fixed parameter,
there is evidence that this parameter is changing over time, with improve-
ments in information technology, widespread adoption of credit scoring
models, and the growth of “infomediaries” such as Dun and Bradstreet.
Petersen and Rajan ~2002! document that the distance between small firms
and their bankers has been growing—from an average of 16 miles in the
1970s to 68 miles in the 1990s—a pattern which they interpret as evidence
that an increasing amount of hard information is being brought to bear on
credit decisions. If they are correct, and if this trend continues, then the
model suggests that the comparative advantage of small banks in small-
business lending may diminish in the future.
24
It should be noted that the implications of the theory for the multibank-holding-company
variable are ambiguous. Perhaps it proxies for complexity and the number of layers of man-
agement, which would tend to discourage small-business lending. But as argued in Section II.A,
a holding-company structure may also make it harder to move capital across divisions, which
would help soft-information research incentives and small-business lending.
25
More precisely, standard measures of borrower credit quality do a better job of explaining
~in an R2 sense! the loan approval decisions of large banks.

Information Production and Capital Allocation
1915
IV. Related Theoretical Work
The ideas in this paper are related to several strands of earlier theoretical
work. Rather than attempting a comprehensive survey, I will just briefly
discuss a few of the most direct linkages. One branch of the literature takes
the perspective that firms are organized so as to be maximally efficient at
the processing and communication of various types of information ~see,
e.g., Sah and Stiglitz ~1986!, Radner ~1993!, Bolton and Dewatripont ~1994!,
Harris and Raviv ~1999!!. Although information production and transmis-
sion are central to my story as well, I differ from these other works in a
couple of ways. First, I focus on managers’ incentives to create and pass on
various types of information, whereas the above-mentioned papers abstract
from agency problems within the firm. Second, the notion of authority is
more prominent in my model. In particular, the CEO does more than just
listen to and act on reports from her subordinates; she actually controls the
resources that these subordinates work with.
The soft-information variant of the model—which emphasizes how the CEO’s
capital-allocation authority can discourage division managers from doing
research—is, as has already been noted, closely related to Aghion and Tirole
~1997!. On a similar note, Rotemberg and Saloner ~1994! argue that firms
may wish to avoid being too broad in scope. For if there are credit con-
straints at the firm level, such narrowness can help the CEO commit to
employees that she will adopt any good ideas that they generate, thereby
strengthening ex ante research incentives. More generally, the idea that agents’
incentives are weaker when they do not have control over asset-allocation
decisions is familiar from the work of Grossman and Hart ~1986!, Hart and
Moore ~1990!, and Hart ~1995!.
However, a sharp distinction between my model and these “costs-of-a boss”
theories arises when information is hard, rather than soft. With completely
hard information, there is no downside to integration in my model. To the
contrary, the fact that division managers do not have control actually serves
to heighten their incentives, as they struggle to produce enough positive
information to convince the CEO to give them a larger share of the capital
budget. Thus, the model not only paints a generally more favorable picture
of the incentive effects of integration than much of the recent literature, its
comparative statics with respect to the hardness0softness of information also
imply more nuanced empirical implications. The empirical distinctions among
the theories are underscored by the facts from the banking industry: The
other costs-of-a-boss stories cannot easily explain why large banks might be
at more of a disadvantage in small-business lending than in, say, credit-card
or mortgage lending.
In some ways, the hard-information version of the model—with managers
trying to convince the CEO to give them more capital—is also reminiscent of
the influence-cost literature ~Milgrom ~1988!, Milgrom and Roberts ~1988!,
Meyer, Milgrom, and Roberts ~1992!!. However, when information is in-
nately hard ~i.e., when b ϭ 1!, the welfare implications are reversed. Unlike

1916
The Journal of Finance
in the influence-cost models, division managers’ efforts to sway the CEO are
productive here, rather than wasteful.26 But this positive result is sensitive
to the details of the information structure. When the hardness of information
is endogenous and b is low ~as in Section II.C!, division managers’ efforts to
attract more capital can lead to inefficient levels of bureaucracy, a result
very much in the spirit of the influence-cost theories.27
Finally, it is worth touching on the connection between my model and
Hart and Moore ~1999!. In their model, some agents ~specialists! have ideas
about individual assets, while other agents ~coordinators! have ideas about
how to use multiple assets together. These ideas are mutually exclusive, so
that only one agent’s idea can be implemented with a given asset. Moreover,
there is no ex post renegotiation. In this setting, the problem is to allocate
decision rights ex ante in such a way as to make sure that the best ideas get
implemented ex post. In contrast, in my model, the “ideas” of the CEO and
the division managers can be ex post complementary. Specifically, the CEO’s
research can help her decide to allocate a certain amount of capital to a
given division manager, who then draws on his own knowledge to make the
right suballocations to individual projects within the division. Thus, there is
no ex post inefficiency when both the CEO and the division managers have
ideas. Rather, the issue is the ex ante one of creating incentives for them to
generate these ideas in the first place.
V. Conclusions
By way of conclusion, it is useful to point out a limitation of the model
developed above. At a fundamental level, the question being asked is: “What
organizational form—decentralization or hierarchy—does the best job of al-
locating capital to competing investment projects?” But in addressing the
question, the focus has been on information production and transmission
inside firms. The notion that valuable information might also be generated
by outside investors—for example, by traders in the stock market—has been
downplayed.28 In particular, I have been assuming that, because their re-
search incentives are weaker than those of a CEO, outside investors acquire
no information about investment prospects and base their decisions simply
on their initial priors.
Clearly, it would be nice to incorporate a richer and more fully endogenous
view of stock-market information production into the model. There are sev-
eral further issues that this sort of extension might allow one to address. In
particular, the quality of stock-market information may—like the quality of
internally generated information—depend on firm size and scope. On the
26
The positive, information-creating effects of self-interested advocacy have also been em-
phasized by Rotemberg and Saloner ~1995! and Dewatripont and Tirole ~1999!.
27
Other papers that stress the negative aspects of intrafirm struggles for capital include
Rajan, Servaes, and Zingales ~2000! and Scharfstein and Stein ~2000!.
28
This idea is stressed by Holmstrom and Tirole ~1993!, among others.

Information Production and Capital Allocation
1917
one hand, a potential advantage of decentralization is that it leads to stock
prices that are specific to narrow, “pure-play” sets of assets, thereby pro-
viding more precise guidance for investment decisions. On the other hand,
pushed too far, decentralization may dampen the overall amount of stock-
market information that is produced: Given fixed costs of information ac-
quisition, tiny firms may not attract much interest from either sophisticated
investors or stock analysts. By taking such factors into consideration, one
might hope to develop a more complete understanding of the link between
organizational form and the efficiency of capital allocation.
Appendix
Proof of Lemma 1: An upper bound on the surplus created by a division
with a three-unit allocation can be obtained by assuming that the division
manager’s research succeeds with certainty, yielding expected net output
equal to ~3g~2! ϩ g~1! ϩ b~2!!04. Conversely, a lower bound for a division with
a two-unit allocation can be obtained by assuming that the division manag-
er’s research never succeeds, in which case expected net output is simply
g~1!. Comparing these two, it must be that two units is preferred if
3~g~2! Ϫ g~1!! ϩ b~2! , 0, which is the sufficient condition given in the
lemma. Q.E.D.
Proof of Lemma 2: An upper bound on the surplus created by a hierarchy
with a five-unit allocation can be obtained by assuming that information is
perfect throughout the hierarchy, so that the five units are always allocated
efficiently. In this case, expected net output is ~7g~1! ϩ 25g~2! ϩ b~2!!016.
Conversely, a lower bound for a hierarchy with a four-unit allocation can be
obtained by assuming that all agents are uninformed, so that each project
always gets one unit of capital, leading to expected net output of 2g~1!. Com-
paring these two, it must be that four units is preferred if 25~g~2! Ϫ g~1!! ϩ
b~2! , 0, which is the sufficient condition given in the lemma. Similar logic
also implies that four units are preferred to six or seven, given the same
condition. Q.E.D.
Proof of Proposition 2: First, straightforward calculation establishes that
a sufficient condition for the CEO to prefer an allocation of at least three
units to a lone-star division ~rather than just giving two units to each divi-
sion! is g~2!02 Ͼ 2g~1!03. This condition is satisfied if, as the proposition
requires, g~2!02 Ͼ 3g~1!04. So it must be that a lone-star division will get at
least three units of funding.
Now suppose that a lone-star division does in fact get exactly three units
of funding. It is easy to show that equation ~9! in the text is modified so
that the utility gain ⌬h4s when a division manager’s research is successful is
given by
⌬
hs ϭ ⌬d4 Ϫ q~2g~2! Ϫ 3g~1!!016.
~A1!
4

1918
The Journal of Finance
But given the condition in the proposition, we know that ~2g~2! Ϫ 3g~1!! is
positive. So ⌬h4s Ͻ ⌬d4, which, in turn, implies that e4hs Ͻ e4d. Thus, even when
the tilt in the capital budget is less extreme, a hierarchy induces less re-
search effort than under decentralization. That a hierarchy also may lead to
lower output—that is, that we may have Y hs~4! Ͻ Y d~4!—follows from the
same reasoning as in the example that was used to illustrate Proposition 1
in the text. Q.E.D.
Proof of Proposition 3: I begin by assuming that division-manager report-
ing strategies are as described in part ~a! of the proposition; I will verify
momentarily that these strategies are optimal for the division managers.
With these reporting strategies, the CEO updates on a quiet division man-
ager using equations ~12!–~14! in the text. Given these updating rules, let us
ask what the CEO does when she faces one division manager that reports
$G, G% and another one that is quiet. It is not hard to show that, for any
value of zpc, it will be optimal for the CEO to give at least three units of
capital to the $G, G% division if the following sufficient condition ~which is
the one required in the statement of the proposition! holds: g~2!02 Ͼ 3g~1!04.
Conversely, it is easy to see that there is no advantage to deviating from the
equal-funding allocation under any other circumstances.
If the CEO follows these capital-allocation rules, then it must in fact be
optimal for a division manager to speak up when his information is either
$G, G% or $G, B%. In the former case, he may get a third or fourth unit of
capital by speaking up, and, in the latter case, he ensures that he will not
be reduced down to one or zero units ~which could happen if he were quiet
and the other division reported $G, G%!. It is also at least weakly better for
the division manager to remain silent when his information is $B, B%. In-
deed, for many parameter values, it is strictly better, since a division that
reports $B, B% will get allocated zero units of capital in circumstances when
a quiet division would get one unit.
These arguments establish parts ~a! and ~b! of the proposition. To prove
part ~c!, assume that when a $G, G% division manager is paired with a quiet
one, the former is allocated exactly three units of capital. This will establish
a lower bound on the gains to being informed, as being informed would be
strictly more attractive if a $G, G% division were to get four units of capital
in this situation. Now take the perspective of manager i, assuming that
manager j exerts effort of ej and therefore has a probability of research
success of p~ej!. ~This implies that there is a probability zp~ej! that man-
ager j will uncover hard information, and a probability zp~ej!04 that man-
ager j will be able to document to the CEO that his division is $G, G%.!
hh
4i
A little algebra yields the following expression for the utility gain ⌬ to
manager i if his research is successful:
⌬
hh ϭ ⌬d4 ϩ z04 ϩ z~4g~2! Ϫ 4g~1!!016 ϩ zp~ej !~2g~1! Ϫ 2g~2!!016
4i
~A2!
ϩ z2 p~ej !~2g~1!!016.

Information Production and Capital Allocation
1919
This implies that
hh Ͼ ⌬d4 ϩ z04 ϩ zp~ej !~2g~2! Ϫ 2g~1!!016 ϩ z2 p~ej !~2g~1!!016.
⌬
~A3!
4i
hh
It then follows immediately from ~A3! that ⌬ Ͼ ⌬d4 ϩ z04 for any value of
4i
p~ej!, which establishes part ~c! of the proposition. Part ~d! is then obvious,
since in a hierarchy with hard information, there is both more information
produced than under decentralization, plus the added advantage that the
CEO reallocates funds across divisions according to a value-maximizing
criterion. Q.E.D.
Proof of Proposition 4 (Sketch): There are two cases to consider. In the
first, the credit constraint facing a decentralized division remains at two
units, but that for a hierarchy is relaxed, and exceeds four units. In this
case, revealed preference on the part of outside investors implies that the
hierarchy must do at least as well as it did with four units under the con-
ditions of Proposition 3, so it continues to dominate decentralization.
In the second case, suppose that each decentralized division can raise three
units. In this case, a hierarchy must be able to raise at least six units, and
maybe more. One can thus put a lower bound on the value created by a
hierarchy by assuming that it raises exactly six units. In a hierarchy with
six units, it is easy to verify that the following constitutes an equilibrium,
for any parameter values: ~a! a division manager speaks up if his informa-
tion is $G, G% and keeps quiet otherwise; ~b! when just a single division
reports $G, G%, the CEO gives it four units of financing; and ~c! in all other
cases, both divisions get three units of financing. It then follows from argu-
ments analogous to those in the proof of Proposition 3 that research incen-
tives are stronger in a hierarchy, and, that as a result, net output is also
greater in a hierarchy. Q.E.D.
REFERENCES
Aghion, Philippe, and Jean Tirole, 1997, Formal and real authority in organizations, Journal of
Political Economy 105, 1–29.
Antle, Rick, and Gary D. Eppen, 1985, Capital rationing and organizational slack in capital
budgeting, Management Science 31, 163–174.
Baker, George, Robert Gibbons, and Kevin J. Murphy, 1999, Informal authority in organiza-
tions, Journal of Law, Economics, and Organizations 15, 56–73.
Berger, Allen N., Rebecca S. Demsetz, and Philip E. Strahan, 1999, The consolidation of the
financial services industry: Causes, consequences, and implications for the future, Journal
of Banking and Finance 23, 135–194.
Berger, Allen N., Anil K. Kashyap, and Joseph M. Scalise, 1995, The transformation of the U.S.
banking industry: What a long, strange trip it’s been, Brookings Papers on Economic Ac-
tivity, 55–218.
Berger, Allen N., Anthony Saunders, Joseph M. Scalise, and Gregory F. Udell, 1998, The effects
of bank mergers and acquisitions on small business lending, Journal of Financial Eco-
nomics 50, 187–229.
Berger, Allen N., and Gregory F. Udell, 1996, Universal banking and the future of small busi-
ness lending, in A. Saunders and I. Walter, eds.: Financial System Design: The Case for
Universal Banking ~Irwin, Burr Ridge, IL!.

1920
The Journal of Finance
Bolton, Patrick, and Mathias Dewatripont, 1994, The firm as a communication network, Quar-
terly Journal of Economics 109, 809–839.
Brickley, James A., James S. Linck, and Clifford W. Smith, 2000, Boundaries of the firm: Evi-
dence from the banking industry, Working paper, Rochester University.
Cole, Rebel A., Lawrence G. Goldberg, and Lawrence J. White, 1997, Cookie-cutter vs. charac-
ter: The microstructure of small business lending by large and small banks, Working paper,
New York University.
Cremer, Jacques, 1995, Arm’s length relationships, Quarterly Journal of Economics 110, 275–295.
Demski, Joel S., 1994, Managerial Uses of Accounting Information ~Kluwer Publishers, Boston,
MA!.
Dewatripont, Mathias, and Jean Tirole, 1999, Advocates, Journal of Political Economy 107,
1–39.
DeYoung, Robert, Lawrence G. Goldberg, and Lawrence J. White, 1997, Youth, adolescence and
maturity of banks: Credit availability to small business in an era of banking consolidation,
Working paper, New York University.
Donaldson, Gordon, 1984, Managing Corporate Wealth ~Praeger Publishers, New York!.
Dye, Ronald A., 1985, Disclosure of nonproprietary information, Journal of Accounting Re-
search 23, 123–145.
Gertner, Robert H., David S. Scharfstein, and Jeremy C. Stein, 1994, Internal versus external
capital markets, Quarterly Journal of Economics 109, 1211–1230.
Grossman, Sanford J., and Oliver D. Hart, 1986, The costs and benefits of ownership: A theory
of vertical and lateral integration, Journal of Political Economy 94, 691–719.
Harris, Milton, and Artur Raviv, 1996, The capital budgeting process, incentives, and informa-
tion, Journal of Finance 51, 1139–1174.
Harris, Milton, and Artur Raviv, 1999, Organization design, Working paper, University of Chicago.
Hart, Oliver D., 1995, Firms, Contracts, and Financial Structure ~Oxford University Press,
Oxford!.
Hart, Oliver D., and John Moore, 1990, Property rights and the nature of the firm, Journal
of Political Economy 98, 1119–1158.
Hart, Oliver D., and John Moore, 1999, On the design of hierarchies: Coordination versus
specialization, Working paper, Harvard University.
Holmstrom, Bengt, and Jean Tirole, 1993, Market liquidity and performance monitoring, Jour-
nal of Political Economy 101, 678–709.
Ijiri, Yuji, 1975, Theory of Accounting Measurement ~American Accounting Association, Chicago,
IL!.
Keeton, William R., 1995, Multi-office bank lending to small businesses: Some new evidence,
Federal Reserve Bank of Kansas City Economic Review 81, 63–75.
Meyer, Margaret, Paul R. Milgrom, and John Roberts, 1992, Organizational prospects, influ-
ence costs, and ownership changes, Journal of Economics and Management Strategy 1,
9–35.
Milgrom, Paul R., 1988, Employment contracts, influence activities, and efficient organiza-
tional design, Journal of Political Economy 96, 42–60.
Milgrom, Paul R., and John Roberts, 1988, An economic approach to influence activities in
organizations, American Journal of Sociology 94, Supplement, 154–179.
Nakamura, Leonard I., 1994, Small borrowers and the survival of the small bank: Is mouse
bank mighty or Mickey? Federal Reserve Bank of Philadelphia Business Review, November0
December, 3–15.
Peek, Joe, and Eric S. Rosengren, 1996, Small business credit availability: How important is
size of lender?, in A. Saunders and I. Walter, eds.: Financial System Design: The Case for
Universal Banking ~Irwin, Burr Ridge, IL!.
Peek, Joe, and Eric S. Rosengren, 1998, Bank consolidation and small business lending: It’s not
just bank size that matters, Journal of Banking and Finance 22, 799–819.
Petersen, Mitchell A., and Raghuram Rajan, 2002, Does distance still matter? The information
revolution in small business lending, Journal of Finance, forthcoming.