Investor Protection, Ownership, and the Cost of Capital

Himmelberg, Hubbard, and Love combine the agency theory of the firm with risk diversification incentives for insiders. Principal-agent problems between insiders and outsiders force insiders to retain a larger share in their firm than they would under a perfect risk diversification strategy. The authors predict that this higher share of insider ownership and the resulting exposure of insiders to higher idiosyncratic risk will result in underinvestment and higher cost of capital. Using firm-level data from 38 countries, the authors provide evidence in support of their theoretical model, showing that the premium for bearing idiosyncratic risk varies between zero and six percent and decreases in the level of outside investor protection. The results of the study imply that policies aimed at strengthening investor protection laws and their enforcement will improve capital allocation and result in higher growth. This paper - a product of Finance, Development Research Group - is part of a larger effort in the group to study corporate governance and access to finance.


Introduction
In this paper, we investigate the e¤ect of investor protection on the cost of capital, where "investor protection" refers collectively to those features of the legal, institutional, and regulatory environment -and characteristics of …rms or projects -that facilitate …nancial contracting between inside owners (managers) and outside investors. Building on the agency framework of Jensen and Meckling (1976) and ideas from the law and …nance literature (e.g., La Porta, Lopez-de-Silanes, Shleifer, and Vishny -"LLSV", 1998), we investigate the empirical implications of investor protection using structural equations derived from a model of inside ownership and investment. In the model, insiders can divert value (or "steal") from outside investors at a cost which depends on the exogenous level of investor protection and the endogenous fraction of equity owned by insiders. Endogenous ownership incentives are expensive to provide, however, for the familiar reason that insiders are forced to bear undiversi…ed idiosyncratic risk. If the exogenous level of investor protection were perfect, insiders would optimally choose to sell 100% of the equity (to diversify fully idiosyncratic risk) and steal nothing, but with imperfect investor protection, this contract cannot be (costlessly) enforced. By retaining a higher fraction of equity, insiders can credibly commit to lower rates of stealing, but are forced to bear higher levels of diversi…able risk.
The tradeo¤ between risk and incentives distorts insiders' incentive to invest in risky capital projects, even under the optimal ownership structure. This is because the cost of capital includes an additional premium for holding idiosyncratic risk which is absent when investor protection allows insiders to diversify fully. Thus the model determines not only the endogenous structure of ownership structure but also the endogenously determined cost of capital and level of capital investment. Our empirical strategy exploits the equilibrium relationship between inside ownership and the marginal return on capital implied by the model. In countries like the United States where investor protection is high, the model predicts endogenously low levels of insider ownership. Accordingly, the idiosyncratic risk premium applied to the cost of capital is low, and the steady-state level of capital approaches the …rst best level of e¢ciency that would obtain in the absence of …nancial contracting costs.
In countries like Turkey or Peru, however, where investor protections are ostensibly weaker, the optimal ownership structure obliges insiders to hold large equity stakes and therefore bear large amounts of idiosyncratic risk, which implies steady-state levels of capital below the …rst-best level.
While the model helps to formalize our intuition, it more importantly formalizes the empirical speci…cation used to investigate the predicted relationship among investor protection, inside ownership concentration, and the cost of capital. Using …rm-level data from Worldscope for 38 countries, we investigate two predictions. First, we estimate the determinants of the fraction of equity owned by insiders. We verify that, as predicted, this fraction depends on measures of investor protection. We emphasize that investor protection has an important cross-…rm dimension in addition to its more familiar cross-country dimension.
Assets like factories that are di¢cult to steal provide a built-in degree of investor protection, whereas assets like the insiders' accumulated knowledge of the product market may be easier to expropriate if these employees can leave to start their own …rms. 1 This cross-…rm variation of investor protection can also explain the cross-sectional di¤erences in the level of inside ownership observed, say, within the United States.
Second, and more important, we document a positive correlation between inside equity ownership and the marginal return to capital, a relationship which follows directly from the …rst-order condition for capital. The cost of capital in the …rst-order condition capital includes a risk premium that re ‡ects the insiders' exposure to idiosyncratic risk. The higher the equilibrium level of inside ownership, the higher the risk premium in the marginal cost of capital. This explains the positive relationship between the marginal return to capital and inside ownership. In addition to providing a test of the above qualitative prediction, this equation allows us to obtain estimates of the steady-state risk premium. We estimate average premiums in the range of zero to …ve percent. Incorporating this value into the model and using the observed levels of inside ownership allows us to assess the magnitude of the capital distortions implied by weak investor protection. Though we consider these estimates and calculations exploratory, they imply that capital stock levels in countries with weak investor protections are less than half the level implied for countries like the United States and the United Kingdom.

Related Research
The research agenda that began with the pioneering work of Alchian and Demsetz (1972) and Jensen and Meckling (1976) has …rmly established agency theory as a basic building block of corporate …nance, but there have few attempts to integrate production theory with the agency theory of corporate …nancial behavior in a uni…ed model of the …rm suitable for structural empirical estimation. In this paper we derive a simple empirical model that builds on the recent work of Burkart, Gromb, and Panunzi (1997), LLSV (1998LLSV ( , 1999, Shleifer and Wolfenzon (2000). Like these papers, our goal is to understand the e¤ect of investor protection on real and …nancial behavior. We borrow from these papers the assumption that "investor protection" can be modeled as a parameter in a cost-of-stealing technology that makes it costly (to varying degrees) for insiders with control over the …rm's decision-making process to "steal" from outside (minority) shareholders. In contrast to the above models, we interpret investor protection as a parameter that varies not only across countries but also across …rms. Consistent with standard agency models, but in further contrast to the previous research, we introduce insider risk aversion as the o¤setting cost of insider ownership. Integrating this agency model of ownership with a conventional production technology generates the basic insight for the cost of capital, and our emphasis on this dimension of the problem is the primary distinguishing characteristic of our paper.
In further contrast to previous research, we use the model to derive and estimate structural equations that we use to help understand the implications of unobserved heterogeneity resulting from the econometrician's incomplete measurement of investor protection. We also use the model to estimate the size of the additional risk premium in the marginal cost of capital, and use this to calculate the magnitude of investment distortions at the …rm level.
There is a large literature recently surveyed by Hubbard (1998) which examines the extent to which investment decisions are a¤ected by …nancial frictions. A recent paper by Demirgüc-Kunt and Maksimovic (1999) investigates whether such frictions are related to country-level measures of …nancial development and investor protection. They …nd that the fraction of …rms growing faster than a "benchmark" model of unconstrained growth is positively related to indicators of …nancial development. Love (2001) estimates Euler equations and similarly …nds that the marginal cost of funds also depends on countrylevel measures of investor protection. Both of these papers recognize the importance of using model structure to control for investment opportunities, but like previous research (Whited, 1992;Gilchrist and Himmelberg, 1998), such models are truly structural only under the null hypothesis of frictionless capital markets. Under the alternative hypothesis of …nancial frictions, the "…nancial side" of such models generally consists of little more than ad hoc model assumptions such as, for example, that the cost of capital is increasing in leverage and dividends are constrained to be non-negative. In this paper, by contrast, ownership structure and leverage are endogenous, and the additional "wedge" for external equity derives from the underlying agency costs. Moreover, the magnitude of this wedge is endogenously re ‡ected by ownership structure. This result follows directly from the …rstorder conditions of a simple model and represents an empirical prediction which previous work has apparently not explored, namely, the predicted relationship between the marginal pro…t of capital and inside ownership.
Our framework sheds light on the structural interpretation of "ownership-performance" regressions of the sort estimated, for example, by Demsetz and Lehn (1985), MÁrck, Shleifer, 5 and Vishny (1988), McConnell and Servaes (1990), Himmelberg, Hubbard, and Palia (1999), and Holderness, Kroszner, and Sheehan (1999). Our model suggest interpretations of these regression results which di¤er sharply from those that have been suggested in past work (including our own), and more generally highlight the dangers of failing to recognize fully the joint endogeneity of ownership variables and balance sheet ratios.
Our focus on the relationship between investor protection and the cost of capital complements research which has attempted to determine whether cross-country variation in …nancial development is associated with investment and growth rates across countries, industries, and …rms. A large body of research documents a link between …nancial development and economic growth using aggregate data (King and Levine, 1993;Levine and Zervos, 1998;Rousseau and Wachtel, 1998;Demirgüc-Kunt and Maksimovic, 1998;and Beck, Levine, and Loyaza, 2000). Rajan and Zingales (1998) use industry growth in the United States as a proxy for the investment opportunities of similar industries outside the United States to show that industries in countries with lower levels of …nancial development grow at slower rates. Consistent with these results, Wurgler (2000) uses industry data to show that the sensitivity of investment growth to value added growth (i.e., investment opportunities) is lower in countries with poorly developed …nancial markets.
The results in this paper are also related (though less directly) to research which seeks to understand the role of ownership rights for investor protection (Grossman and Hart, 1988;Stulz, 1988;Zwiebel, 1996;Fluck, 1998;and Myers, 2000). 2 For example, one-share, one-vote rules -which are often cited as being good for investor protection -have been analyzed in detail by Grossman and Hart (1988). The cost-of-stealing model used here does not explicitly model control rights, but we nevertheless view control considerations as an important determinant of the exogenous level of investor protection. 3 Finally, there is 2 Identifying the sources of investor protection is one of the primary questions in the research on corporate governance recently surveyed by Shleifer and Vishny (1997). 3 Burkart, Gromb, and Panunzi (1997) consider a model in which both cash ‡ow and control rights are used to provide incentives for managers. They point out that the free-rider problem by target shareholders limits 6 a related literature which emphasizes the role of the legal system for investor protection. Levine (1999), for example, argues that the legal system is a key determinant of both …nancial development and economic growth, while LLSV (1997LLSV ( , 1998 and Co¤ee (2000) argue that common law systems provide stronger investor protection than civil law systems.
The empirical model in this paper does not attempt to formalize the workings of alternative legal regimes; this is well beyond our scope. Instead, we summarize the e¤ect of the legal system by positing an empirical mapping from observable features of the legal environment into a single parameter indexing the "cost of stealing," i.e., the level of investor protection.
As we show, this characterization of the contracting environment does not necessarily limit our ability to assess many qualitative and quantitative implications of the model. The remainder of the paper is organized as follows. We begin in section 2 by introducing a simple model from which we derive implications for ownership and the cost of capital.
Section 3 explores econometric issues that arise in the speci…cation of the empirical model, followed by empirical results in section 4. Section 5 discusses some interesting implications and applications, and section 6 concludes.

The Model
Consider the two-period problem confronting an entrepreneur (alternately "manager" or "insider") who is initially endowed with liquid wealth W it and a project which yields a total return of ¦ (K it ; µ it ), where K it denotes the stock of …xed capital. For simplicity, we assume there are no adjustment costs, and normalize the purchase price of capital to one; we add adjustment costs later in the empirical speci…cation. In the …rst period, the entrepreneur can sell equity or borrow to …nance capital expenditures, K it , and consumption, C it . Equity …nancing X it is raised by selling claims to a fraction 1 ¡ ® it of future dividends. Borrowing the incentive properties of disciplinary takeovers, and therefore tends to favor using cash ‡ow rights to align insider incentives. Their results provide some justi…cation for our simpli…ed model in which the allocation of cash ‡ow rights is endogenous, but the allocation of control rights is captured by the cost-of-stealing function. 7 (or saving) occurs at the rate r t+1 . The borrowing-saving rate need not be riskless (e.g., the manager can invest in the market portfolio), but we assume the return cannot be made contingent on the idiosyncratic outcome of the …rm. This assumption is important, and is meant to capture the intuition that equity (and not debt) is the natural instrument for sharing the …rm's idiosyncratic risk.
The agency problem between insiders and outsiders arises because insiders can steal or divert a fraction s it+1 of …rm pro…ts to themselves before paying dividends. The manager cannot costlessly commit in period one to the level of stealing in period two. Stealing is, however, discouraged by an exogenous punishment technology which imposes a monetary cost c (Á it ; s it ) = 1 2 Á it s 2 it . The parameter Á it is therefore a quantitative index of investor protection, where higher parameter values impose a higher cost of stealing, and therefore indicate better protection. The parameter Á it is easy to interpret because it is proportional to the cost of stealing; to double the cost of stealing, for example, we double Á it : 4 Under this functional form assumption, the total and marginal costs of stealing are increasing in Á it , so that c Á > 0 and c sÁ > 0. This functional form also has the intuitively appealing property that the cost of stealing be convex in s it . According to the model, "investor protection" is anything that exogenously increases the cost to insiders of stealing from outsiders. In particular, the model does not distinguish …rm-level and country-level determinants of investor protection; the parameter Á it is meant to summarize the net impact of all features of the contracting environment. Thus a …rm operating hard-to-steal assets in a country with weak legal enforcement could have insider ownership levels comparable to a …rm operating easy-to-steal assets in a country with strong legal enforcement. 5 Our empirical speci…cation for inside equity ownership explicitly allows 4 If the probability of disciplinary takeover were a cost of stealing, and if voting rights were exogenously tied to dividend rights, one could argue that the takeover probability is an increasing function of 1 ¡´i®it, where´i¸0 is an inverse index of the e¤ectiveness of the market for corporate control. This would imply a generalized stealing function of the form c (Á i ; sit; ®it) = 1 2 Á i s 2 it (1 ¡´i®it). We assume the probability disciplinary takeovers does not much depend on the allocation of dividend rights (that is, Á it > 0 and´i = 0). Arguments in Burkart, Gromb, and Panunzi (1997) are consistent with this assumption. 5 Another …rm-level characteristic on which investor protection might also depend is the identity of the 8 for such cases.
To the extent that insiders own equity in the …rm, they only steal from themselves. Inside ownership of equity therefore provides a mechanism with which managers can commit to lower levels of future stealing. Under the above assumptions, stealing at the rate s it generates a direct bene…t of (s it ¡ c (Á it ; s it )) ¦ (K it ; µ it ) for insiders, and leaves (1 ¡ s it ) ¦ (K it ; µ it ) to be divided up among shareholders (including the inside shareholders). The manager's net return N it+1 in period t + 1 from operating the …rm is therefore: Equity proceeds raised from outside investors must guarantee (in expectation) the market rate of return. If investors value next-period cash ‡ows according to the stochastic discount factor M t+1 , the proceeds from selling a fraction 1 ¡ ® it of the equity is given by: Stealing occurs in the second period after the proceeds X it have been raised. Thus the second-period level of stealing maximizes equation (1) without regard for equation (2), and is characterized by the …rst-order condition where c s (Á it ; s it+1 ) denotes the derivative of c with respect to s it+1 . This equation says that at the optimum, the marginal cost of stealing, c s (Á it ; s it+1 ), plus the marginal reduction of the insiders' dividends, ® it , is equated with the marginal bene…t of stealing, which equals one. If the cost-of-stealing function is monotonically increasing, as we assume, then stealing minority shareholders. For example, foreign investors may be treated di¤erently than domestic investors if they carry less political clout with law enforcement agencies.
is monotonically increasing in outside ownership. As in LLSV (1999a), we assume the functional form c (Á it ; s it+1 ) = 1 2 Á it s 2 it+1 , in which case optimal stealing is given by In the language of principal-agent theory, equation (3) represents the manager's incentivecompatibility constraint, and equation (2) represents the investors' participation constraint.
Both of these constraints must be recognized by the managers and investors in period one when the choices of ® it , K it+1 ; and C it are made. The manager's problem is therefore to choose the vector f® it ; s it+1 ; K it+1 ; C it g to maximize total expected utility, subject to equations (1), (2), and (3) and the budget constraint, given by where A it = W it + X it ¡ K it+1 ¡ C it is the manager's net position in the market asset.
We do not impose any constraints or penalties on the amount of saving or (default-free) borrowing. It is often argued that debt helps to reduce agency costs because it represents a harder claim which reduces the free cash ‡ows from which managers can steal. We assume debt is repaid with probability one. In other words, managers can credibility promise not to steal from debt holders and therefore riskless debt is frictionless. In this sense, managers are not borrowing-constrained -debt markets are willing to let managers borrow as much as they need. Despite this, managers have strong ex ante incentives to use outside equity because default-free debt cannot be used to diversify the idiosyncratic risk. Thus the burden of risk sharing falls solely on equity. At the cost of additional complexity, our model could be generalized to allow risky debt. This might yield additional interesting predictions for leverage, but because debt is such a crude instrument for risk sharing, we think it is unlikely that this would substantially change the qualitative or quantitative predictions of the model for ownership.

The Benchmark Case: Perfect Investor Protection
If the manager could contractually commit to the level of stealing in period two (i.e., if investor protection were "perfect", so that Á it = 1), it follows immediately from equation (3) that regardless the level of managerial ownership, the manager would optimally choose to steal nothing. In this case there is no incentive bene…t from having the managers retain an equity stake in the …rm, so diversi…cation motives make it optimal to sell 100% of the equity to outside investors. It is easy to show in this case that the …rst-order condition for capital is: where ¦ K it+1 = @¦ it+1 =@K it+1 is the marginal value of capital. This is the standard …rstorder condition for the e¢cient choice of capital. To put this equation in more familiar terms, denote the total return on capital by ¦ it = ¼ it + (1 ¡ ±) K it , where ¼ it denotes the current level of variable pro…t, ± denotes the rate of physical depreciation on capital, and (1 ¡ ±) K it represents the resale value of the capital stock (we maintain the assumption of zero adjustment costs). By assumption, the market's stochastic discount factor (SDF) , where r f t+1 is the risk-free rate. Hence we can write the previous equation as: where ¼ K it+1 = @¼ it+1 =@K it+1 is the marginal pro…t of capital. The right-hand side of this equation represents the …rm's "user cost of capital," which is the sum of the (risk-adjusted) opportunity cost of funds and depreciation costs. The covariance between the market's SDF and the marginal pro…t of capital (scaled by E t [M t+1 ]) is non-zero to the extent that …rm pro…ts are a¤ected by (nondiversi…able) aggregate shocks. For example, if ¼ K it+1 were negatively correlated with the market's SDF (i.e., if the …rm had a positive "beta"), its payout would on average be high in states of the world where high payouts are valued less.
< 0 would imply a positive risk premium. As usual, idiosyncratic shocks to ¼ K it+1 (i.e., shocks that are orthogonal to M t+1 ) are not priced because it is assumed they can be costlessly diversi…ed by outside investors. In short, our discussion thus far has produced the textbook advice for managers: Invest up to the point where the expected marginal pro…t of capital equals the user cost of capital, where the user cost is adjusted for nondiversi…able risks (and ignores idiosyncratic risk).

Imperfect Investor Protection
When investor protection is not perfect (that is, when exogenous costs of stealing are not in…nite, or Á it < 1), agency con ‡icts arise. Such contracting frictions could arise for a variety of reasons. For example, it could simply be the case that stealing is unobservable.
Even if stealing is observable, however, frictions could still arise because contract enforcement is costly and unreliable. Although the former interpretation is common in the classical analysis of agency problems, the latter interpretation is a better description of the empirical setting we have in mind. It easily accommodates interpretations based on the quality of the exogenous contracting environment as determined by the legal system such as, for example, laws or judicial traditions which determine the protection of minority shareholders. Such protections are summarized by the cost-of-stealing parameter, Á it .
It is straightforward to show that the …rst-order condition characterizing the optimal capital choice is: is the SDF for the manager. 6 To simplify notation, equation (8) uses: where s it+1 denotes the optimal (ex-post) level of stealing, which is itself a function: Note the contrast between m it+1 , which is the SDF for the manager, and M t+1 , which is the SDF for the market. Under complete markets (complete risk-sharing), the covariance properties of M it and m it are the same. In the current setting, however, risk sharing is incomplete due to the existence of moral hazard, and the covariance properties of M it and m it are not the same.
The …rst-order condition for capital can alternatively be written: 7 This equation says the risk adjustment to the user cost of capital is the weighted sum of two terms. The …rst term, , re ‡ects the covariance between the manager's SDF and the marginal pro…t of capital. To the extent that a sizeable fraction of the manager's income is derived from the pro…tability of the …rm, the manager's consumption is exposed to idiosyncratic risk. In particular, idiosyncratic pro…t shocks increase ¼ K it+1 and consumption, thus decreasing the marginal utility of consumption, which implies cov t £ The second term, , re ‡ects the usual compensation for nondiversi…able risk (just as in equation (7)). When the equilibrium level of stealing is "small," then g it and h it approximately equal ® it and 1 ¡ ® it , respectively. Thus the fraction of equity held by managers reveals the extent to which the user cost of capital applied by the managers re ‡ects idiosyncratic as opposed to systematic risk. When ® it = 0, outside investors own all of the equity in which case only the systematic risk of the …rm is priced. At the other extreme, when ® it = 1, the …rm is a proprietorship and the total risk of the …rm is priced according to the manager's SDF.
Additional structure on the nature of the above risk premiums is provided by the insiders' ownership choice. The …rst-order condition for ownership implies: Under our functional form assumptions on the cost of stealing, g ® it = 1 ¡ s it and h ® it = 2s it ¡ 1. Hence equation (13) can be re-written as: which implies: This equation says that managers assign a lower value to risky pro…ts than outside investors do. If investor protection were perfect, the level of stealing would be zero, and these values would be equal. Under imperfect investor protection, however, managers assign a lower value to stochastic pro…ts because they discount for idiosyncratic risk, whereas the market, by contrast, is indi¤erent to this risk. The manager's ownership choice is nevertheless privately optimal because the marginal value of reducing idiosyncratic risk exposure by selling more equity equals the marginal reduction in the market price this would require in compensation for the higher rate of equilibrium stealing that would accompany the lower ownership stake.
If we assume the value function ¦ it+1 is homogenous of degree one in the capital stock, equation (13) can also be used with equation (8) to derive an alternative expression for the …rst-order condition for capital. Linear homogeneity implies ¦ it+1 = K it+1 ¦ K it+1 , which allows us to combine equations (8) and (13) to obtain: Under our functional form assumption for the cost of stealing, , hence we can rewrite this equation as: It follows immediately, that E t [M t+1 ¦ it+1 ] > 1. From the market's perspective, this equation says that the marginal value of pro…t exceeds its purchase price. That is, in contrast to the benchmark case of perfect investor protection characterized in equation (6), the manager is underinvesting.
The magnitude of the "wedge" between the …rst and second best allocations of capital is roughly proportional to the equilibrium level of stealing, s it . For example, suppose the level of managerial ownership were ® it = 0:4, which is the median in our sample. Suppose further that the equilibrium rate of stealing were a (relatively modest) two percent (s it = 0:02).
Then 1 2 s it (3 + ® it ) = 0:034. That is, such a …rm would invest as if its cost of capital were about three and a half percentage points higher. Increasing the assumed equilibrium level of stealing to …ve percent implies a marginal cost of capital of over eight percentage points higher! Cost of capital di¤erences of this magnitude are large enough to have …rst-order e¤ects on …rm size and the growth and development of industries and countries. This 15 motivates the empirical investigation in the remainder of the paper.

Empirical Implications
The primary goal of our empirical work is to investigate the …rst-order condition for capital in equation (12) or equation (17). In practice, estimation of either equation is complicated by two issues. First, should we assume that the econometrician observes ¼ it+1 ? Or should we recognize that perhaps "after-stealing" pro…ts are being reported, given that we do not observe stealing, how do we evaluate the expressions for g it and h it ?
Regarding the measurement of pro…ts, reasonable arguments can be made both ways depending on whether stealing is deducted from accounting pro…ts. On the one hand, if self-dealing which takes the form of a manager purchasing input goods from a relative at in ‡ated prices, then the econometrician measures (1 ¡ s it ) ¼ it+1 . On the other hand, if self dealing involves stock transactions that bene…t managers at the expense of minority shareholders, then accounting pro…t is correctly measured. As a practical matter, we are inclined to think the former is more descriptive in most settings. In this case, equation (12) can be formulated in terms of observed marginal pro…t as: where:°i We are still not ready to estimate equation (18) because neither s it nor°i t (nor ¡ it ; for that matter) is observable in the data. In particular, one cannot calculate the necessary covariance without observing m it+1 , which requires knowing the current and future values of the manager's consumption.
Our empirical investigation is based on equation (12) and proceeds from the assumption that s it is "small" relative to ® it . This implies g it ' ® it , and h it ' 1 ¡ ® it : Next, we model°i t and ¡ it using variable coe¢cient models in which we assume°i t = ¹°+ "°i t and This allows us to write equation (18) as and ! it is a rational expectations error orthogonal to information at time t: In the presence of the random coe¢cient error, , we need to consider whether an instrumental variable estimator based on instruments z it in the time-t information set satis…es E [u it z it ] = 0. Given the rational expectations error ! it , and given no reason to expect covariation between " ¡ it and ® it , the validity of the moment condition reduces to establishing zero conditional covariance between ® it and "°i t . This condition is not easy to verify a priori because the covariance depends on the source of the underlying shocks. In the model, a negative shock to the insiders' private wealth would imply a negative reponse of "°i t (because the marginal utility of wealth is lower in good states) and a positive response of ® it (because lower risk aversion encourages more inside ownership). That is, wealth shocks would imply E ["°i t ® it jz it ] < 0; which would bias estimates of ¹°¡ ¹ ¡ in equation (21) toward zero. Alternatively, the covariance implied by shocks to investment opportunities would imply a positive correlation and an upward bias. As a practical matter, ownership stakes tend to evolve slowly, so we are not overly concerned about the magnitude of the bias in either direction, especially when equation (21) is estimated using instrumental variables.

Data
Our empirical investigation uses annual …rm-level data from the Worldscope database, which contains information on large, publicly traded …rms, and monthly …rm-level stock price data from Datastream. 8 All countries in the Worldscope database (May 1999 Global Researcher CD) with at least 30 …rms and at least 100 …rm-year observations are included in the sample.
We exclude data from former socialist economies. This results in a sample of 38 countries.
The sample does not include …rms for which the primary industry is either …nancial (onedigit-SIC code of 6) or service-oriented (one-digit-SIC codes of 7 and above). 8 Worldscope attempts to standardize accounting information to improve cross-country comparability. For example, if one company reports sales with included excise taxes and another company excludes taxes, Worldscope corrects this di¤erence and presents both with taxes excluded. This is important for our purposes because sales is the key ingredient in the measure of the marginal product of capital. It is therefore obviously desirable that it have as much cross-country comparability as possible.
We construct a beginning-of-period capital stock variable which is used to construct investment and sales-to-capital ratios as well as our measure of the marginal product of capital (see the next section). The most obvious measure, the lagged end-of-period capital stock, is problematic because mergers, acquisitions, divestitures, and similar events give rise to large, unexplained changes in ratios using capital in the denominator. There is no easy, systematic way of identifying these transactions in the data, and even if we could, throwing them out would substantially reduce sample size, so we calculate beginning-ofperiod capital stock as the current end-of-period stock minus current period gross investment plus depreciation. We also construct …rm-level measures of the variance of idiosyncratic stock returns. We match monthly stock market data from Datastream to estimate the variance of idiosyncratic returns for over 90% of our Worldscope …rm-year observations. In the raw data, there are a few returns which appear to be outliers (e.g., returns below 100%); these are removed by eliminating values for which the absolute value of returns exceeds 100%; this rule deletes fewer than one tenth of one percent of the observations, and estimates are not sensitive to this cuto¤. Our measure of idiosyncratic risk is the variance of the residual from obtained by regressing monthly …rm-level stock returns on the respective country-level measure of the market return (the country-level market index is also obtained from Datastream).
Inside ownership concentration is a key variable for analysis. Though it is less than the ideal measure, we use the Worldscope variable "closely held shares" as our measure of inside ownership. At the country level, we augment these …rm-level data with three indicators of investor protection which we construct by aggregating the indices of "shareholder rights," "creditor rights," and "legal e¢ciency" assembled by LLSV (1998). The "shareholder rights" index measures how strongly the legal system favors minority shareholders against managers or dominant shareholders in the corporate decision-making process. This index is a sum of seven characteristics, each of which is assigned a value of one if the right increases shareholder protection, and zero otherwise. The components of this index are: (1) one share-one vote rule; (2) proxy by mail; (3) shares not blocked before meeting (in some countries, the law requires depositing shares with the company several days prior the shareholder meeting, a practice which prevents shareholders from selling or voting their shares); (4) cumulative voting/proportional representation; (5) oppressed minority rights (the shareholder right to challenge director's decisions in court or force the company to repurchase the shares from minority); (6) preemptive right to new issues (which protects shareholders from dilution); and (7) percentage of share capital required to call an extraordinary shareholder meeting.
The "creditor rights" index measures the rights of senior secured creditors against borrowers in reorganizations and liquidations. This index is a sum of four characteristics. The components of this index are: (1) no automatic stay on assets (which makes it harder for secured creditors to seize collateral); (2) secured creditors paid …rst; (3) restrictions on going into reorganization (equal to one for countries that require creditors' consent to …le for reorganization); (4) management does not stay in reorganization (equal to one if management is replaced at the start of reorganization procedure). Finally, the "legal e¢ciency" index is an assessment of the e¢ciency and integrity of the legal environment as it a¤ects business, particularly foreign …rms. The index is produced by the country-risk rating agency Business International Corporation. The value we use is the average between 1980-1993, scaled from 0 to 10, with lower scores for lower e¢ciency levels.
Finally, we delete observations meeting any of the following criteria: (1) three or fewer years of coverage; (2) zero, negative, or missing values reported for capital expenditures, capital stock (property, plant, and equipment), sales or closely held shares; (3) investmentto-capital ratios greater than 2.5 (which is the upper …rst percentile); (4) sales-to-capital ratios greater than 20 (which is the upper …fth percentile). 9 Table 1 reports the number of …rm-year observations remaining for each country following the application of the above selection criteria, and Table 2 reports summary statistics for these variables across the three samples. Table 1 shows that the number of …rms varies widely across countries. As noted by LLSV (1997), Worldscope's coverage of …rms within countries varies widely from as little as one percent of all listed domestic …rms included (for India) to as many as 82% (for Sweden).
This variation re ‡ects several factors. Some countries are simply larger, and therefore have more …rms. The sample re ‡ects the endogenous decision of …rms to go public or remain private. For example, there are more …rms in countries like the United Kingdom (993 …rms in the full sample) which have strong legal protection for minority shareholders than there are in countries like Germany (375 …rms in the full sample), which has a larger economy but is thought to have weaker shareholder protection. We have fewer observations for countries like India where, despite a large number of public …rms, many …rms are not actively traded, and Worldscope presumably does not bother to collect data for such …rms. To the extent that weak investor protection lowers market liquidity, this presumably weakens the power of our tests by selecting against the very …rms for which the correlation between inside ownership and the marginal return on capital would presumably be strongest.

Measuring the Marginal Pro…t of Capital
Estimation of the model requires a measure of the marginal pro…t of capital. Suppose the …rm's production function is Y it = f (A it ; K it ; Z it ); where A it is a measure of total factor productivity, Y it is output, K it represents the stock of …xed property, plant and equipment, and Z it is a vector variable factor inputs (e.g., materials, energy, unskilled production workers, etc.). Assuming that the …rm faces an inverse demand curve P (Y it ) and variable factor prices w it (in a competitive factor market), the pro…t function is de…ned by By the envelope theorem, the marginal pro…tability of …xed capital, @¼ it =@K it , is where´´(@Y =@P )P=Y < ¡1 is the (…rm-level) price elasticity of demand. If the production function is assumed to be homogeneous of degree , then where P it Y it =K it denotes the sales-to-capital ratio. Thus, up to a scaling factor ¡ 1 +´¡ 1 ¢ , and assuming the book value of capital is a reasonable proxy for replacement value, the marginal pro…t of capital is easily measured using the sales-to-capital ratio.
We allow for the possibility that the scaling factor ¡ 1 +´¡ 1 ¢ may vary across industries. Following Gilchrist and Himmelberg (1998), we construct estimates of ¡ 1 +´¡ 1 ¢ for each industry by assuming that …rms are, on average, near their equilibrium capital stocks.
In steady state, the expected marginal return on capital equals the user cost of capital: where µ j is the industry-speci…c value of ¡ 1 +´¡ 1 ¢ , and where r and ± are the average riskadjusted required return and depreciation rate of capital, respectively. Replacing population moments with sample moments over all …rms and years in industry j, a consistent estimate 22 of µ j is given by:μ We assume r + ± = 0:18 for all industries (results are not sensitive to alternative assumptions). Thus, is our measure of marginal return to capital.

The Determinants of Inside Ownership
Our …rst empirical exercise estimates the e¤ect of …rm-level and country-level measures of investor protection (described above) on inside ownership. In Table 3 we report coe¢cient estimates for …ve alternative speci…cations for the determinates of inside ownership concentration. The …rst three columns use data for the international sample of …rms. To insure robustness of our results to the possibility of selection bias introduced by the idiosyncrasies of the Worldscope data, column (4) reports estimates using the largest 150 …rms in each country. Columns (5) and (6)  The results reported in Table 3 broadly support the proposition that ownership concentration is determined by the level of investor protection. For the sake of comparison with previous work, the speci…cation in column (1) includes only country-level determinants of investor protection. The coe¢cients on both "legal e¢ciency" and "shareholder protection" are negative and precisely estimated, as predicted by theory, while the coe¢cient on "creditor protection" is not statistically di¤erent from zero. These results are consistent with the results found by LLSV (1998). For the sake of comparison with previous work on …rm-level determinants of ownership, column (2) excludes country-level determinants.
Following Himmelberg, Hubbard, and Palia (1999), the speci…cation includes the log of sales, the ratio of sales-to-capital, the ratio of R&D-to-sales, the standard deviation of the idiosyncratic component of stock returns, two-digit (SIC) industry dummies, and country-speci…c year dummies. We also include the dummy variable RDDUM which equals unity 23 if R&D information is reported. This variable provides an additional discrete indicator of R&D intensity because R&D is usually not reported when the amount is negligibly small.
Columns (3) and (4) combine country-level and …rm-level determinants both with and without the stock sigma. Columns (5) and (6) repeat the speci…cation in column (4)  The coe¢cient on the ratio of sales to capital is positive and statistically signi…cant at the one-percent level in all …ve speci…cations. It is traditional in such regressions to interpret the sales-to-capital ratio as a measure of asset tangibility, because high ratios implicitly indicate the presence of intangible assets like …rm-speci…c human capital, technology, or market power. If intangible assets are easier to divert or steal (perhaps because they are di¢cult to observe), then this would explain why sales-to-capital is such a strong, positive predictor of inside ownership. An alternative explanation for the sales-to-capital ratio is that this correlation arises endogenously because the sales-to-capital ratio is closely related to the marginal pro…t of capital, and hence re ‡ects the relationship in equation (21). This model prediction is the primary focus of the next section. The desire to control for tangibility of assets is also part of the motivation for the inclusion of the R&D-to-sales ratio and the R&D dummy. This argument predicts a positive coe¢cient. The R&D variables could also capture idiosyncratic risk which is not measured by the variance of idiosyncratic stock returns (e.g., peso risk), in which case the predicted coe¢cient would be negative. In addition, it is likely that R&D is endogenous -…rms with better investor protection would have an easier time …nancing R&D, in which case R&D, like low inside ownership, would be an endogenous proxy for good investor protection. This, too, would predict a negative coe¢cient. The coe¢cient estimates in Table 3 are more consistent with the view that R&D is a proxy for unmeasured risk or an endogenous indicator of weak investor protection.
The point estimates on our measure of idiosyncratic risk ("stock sigma") are all negative, though only the estimate in column (5) for the non-US/UK …rms is statistically di¤erent from zero. In the model, the ownership choice equates the marginal bene…ts of incentives and risk sharing; idiosyncratic risk makes it costly for insiders to own equity in the …rm. The results in Table 3 are consistent with this prediction of the model. Alternative explanations are possible, however. For example, Demsetz and Lehn (1985) suggest that stock price volatility could also be a proxy for asymmetric information. If ownership concentration were the result of adverse selection, then the predicted coe¢cient on stock sigma would be positive rather than negative. According to this view, the coe¢cient on sigma would be positive, but the estimates in Table 3 are negative, hence the data are more consistent with moral hazard than adverse selection as an explanation for insider ownership concentration.
Of course, these stories are not mutually exclusive; the coe¢cient on sigma could re ‡ect both e¤ects.
In column (3), our preferred speci…cation, the estimated coe¢cients on legal e¢ciency and shareholder protection are all negative and precisely estimated. These results are robust to the exclusion of smaller …rms outside the largest 150 …rms in each country. The negative signs on legal e¢ciency and shareholder protection support the argument in LLSV (1998) that ownership concentration is a substitute for legal institutions as a mechanism for constraining the expropriation of outside equity investors. The economic intuition for the negative coe¢cient on creditor protection in column (3) is less obvious, but still consistent with this view; to the extent that debt …nancing is costlier due to weak creditor protection, …rms may rely more on equity …nancing. Moreover, the coe¢cients on …rm-level variables are robust to the inclusion of country-level variables, and conversely, the coe¢cients on countrylevel variables are not substantively a¤ected by the inclusion of …rm-level variables. Indeed, the incremental adjusted R 2 more than doubles from 0:112 to 0:233 when the speci…cation using only …rm-level variables in column (2) is expanded to include country-level variables in column (3).
Finally, it is interesting to compare the results for the full international sample in column (3) with the samples of in columns (4) and (5). Although there is some overlap in the samples, it is nevertheless reassuring to note that the results for the full international sample are robust across the two subsamples. Table 4 reports the estimated coe¢cient from simple OLS and instrumental variable regressions of the marginal return on investment (¼ it ) on inside ownership concentration -that is, the speci…cation in equation (21), which for ease of reference is reproduced here:

The First-Order Condition for the Capital Stock
These regressions produce estimates of ¡ ¹°¡ ¹ ¡ ¢ , which is the average additional risk premium for bearing idiosyncratic risk (beyond the usual premium ¹ ¡ for bearing systematic risk, which is absorbed in the constant term and therefore not identi…ed in this speci…cation  Tables 4 and 5 (like Table 3) re ‡ect adjustments to account for the potential presence of heteroskedasticity and cross-sectional correlation among observations within a single …rm, and are therefore as conservative as possible. Most of the speci…cations (as indicated) also include industry and time dummies as controls.
In the …rst column of Table 4 The OLS results in columns (1)-(4) of Table 4 indicate positive and statistically signi…-cant estimates of ¡ ¹°¡ ¹ ¡ ¢ ranging from 0:018 to 0:056. We consider three possible reasons why these estimates might be biased. First, as discussed in section 3, inside ownership is endogenous, raising the potential for bias caused by correlation between inside ownership and the error term. However, it is important to be clear about the source of the endogeneity and its implications for the estimation of equation (21). The endogeneity of ® it is not by itself su¢cient to generate the correlation between ® it and the error term that would bias OLS estimates. Indeed, this endogeneity is the very source of the predicted correlation between ® it and the expectation of ¼ it on which our empirical evidence is based. Moreover, the rational expectations error introduced by the di¤erence between the actual and expected value of ¼ it is not known at the time ® it is chosen and is therefore orthogonal to ® it . In short, the model does not imply any obvious economic sources of correlation between inside ownership and the error term.
Because our data provide only relatively crude measures of inside ownership, however, it is likely that the OLS estimates in Table 4 are contaminated by classical measurement error. In the column (5), we reestimate the speci…cation in column (6) Table 4.

Adjustment Costs and Leverage E¤ects
For simplicity, the speci…cation estimated in Table 4 is derived under the assumption of zero adjustment costs and frictionless debt markets. Previous research, however, shows that both adjustment costs and leverage e¤ects are important features of investment behavior (see Gilchrist and Himmelberg, 1998, for a recent treatment). It is therefore important to show that the estimates in Table 4 do not spuriously re ‡ect either of these two features of a more general model. Fortunately, the necessary model extensions can be applied in a straightforward way to equations (13) and (8), and equation (21) can be modi…ed accordingly.
Adjustment costs can be appended to the existing model by recognizing that the total return on capital, under adjustment costs, where c it+1 is the marginal adjustment cost of installing an additional unit of capital. We assume this marginal adjustment cost can be parameterized as To add leverage e¤ects to the model, we …rst note that the model already allows managers to borrow and save freely at the rate r t+1 . To allow for the further possibility that leverage incurs a deadweight loss which is borne by managers, we can make the common and convenient modeling assumption that the borrowing rate r t+1 includes an additional premium which is linearly increasing in the debt-to-asset ratio. In this case, r f t+1 in equation (28) is replaced by r f t+1 +´(B=K) it (see Gilchrist and Himmelberg, 1998, for example).
The empirical speci…cation of the Euler equation can therefore be written: In the absence of adjustment costs for investment and costly debt …nancing, the reduced-form coe¢cients b 1 , b 2 , b 3 , andá re zero, and equation (29) reduces to the static …rst-order condition for capital in equation (8).
We report estimates of the Euler equation in equation (29) in Table 5. These speci…cations are estimated by instrumental variables where the instrument list consists of lags t¡1, t ¡ 2, and t ¡ 3 of all variables appearing in the model speci…cation being estimated. 11 All speci…cations are estimated with country-speci…c year dummies and industry dummies. For the sake of comparison with the estimates in Table 4, columns (1), (2), and (3) of Table 5 report instrumental variable estimates of our modi…ed Euler equation under the assumption that adjustment costs are zero (column (1) repeats column (5) from Table 5 Table 5). In regression results not reported here, we control for size by including the log of sales; this addition does not substantively alter the estimated coe¢cients or standard errors on inside ownership.
Finally, the results for the non-US/UK sample reported in Panel B are qualitatively the same as those in Panel A, except that the coe¢cient estimates for the static model in Panel B tend to be somewhat larger.
Columns (4), (5), and (6) of Table 5 repeat the speci…cations in the …rst three columns allowing for adjustment costs. Here again, we are primarily interested in noting the impact on the estimated coe¢cient on inside ownership. The coe¢cients on ownership in the Euler equation estimates in Panel A are uniformly higher than the estimates for the static speci…cation reported in Table 4 and the …rst three columns of Table 5. For example, in the column (4) of Panel A, the estimated coe¢cient on inside ownership is 0:052 (with a standard error of 0:023), which is larger though less precisely estimate than the estimate of 0:033 (with a standard error of 0:009) reported in the …fth column of Table 5 Table 5, these results appear to indicate that leverage, too, is correlated with the cost of capital used by insiders to discount future cash ‡ows. This is consistent with the leverage e¤ects for investment found by Whited (1992) and Gilchrist and Himmelberg (1998), among others.

The Magnitude of Capital Stock Distortions
The magnitude of the underinvestment implied by our estimates of ¹°¡ ¹ ¡ in Tables 4 and 5 depend on the distortion to the marginal cost of capital, as revealed by the term ¡ ¹°¡ ¹ ¡ ¢ ® it , and the elasticity of the capital stock to the marginal cost of capital. Although it is perhaps di¢cult to judge the value of this elasticity at the level of the macroeconomy, it is not di¢cult to make reasonable assumptions at the …rm level. The elasticity depends on the curvature of the …rm's pro…t function. If the production function is Cobb-Douglas with constant returns to scale, and if the …rm is a price taker in factor and product markets, then the …rm's pro…t function is linear in capital, and …rm size is indeterminate. To generate a concave pro…t function (so that …rm size is bounded), we need to introduce diminishing marginal revenue.
This would be consistent with decreasing returns to scale in production, market power, or both. For simplicity, we assume constant returns to scale and a downward sloping demand curve for output given by P (Y it ) = Y ¡í t , where ¡´is the inverse price elasticity of demand. For this demand curve, the pro…t function function has the form ¼ it = A it K 1¡í t , where A it is a "pro…tability" parameter that embeds productivity levels, factor prices, and parameters of the production and demand functions. In the absence of adjustment costs, equation (21) implies: Equation (30) allows us to examine the sensitivity of the capital stock to changes in the user cost of capital. When investor protection is perfect, equation (30) implies (1 ¡´) A it K ¡í t = ± + r. Abstracting from adjustment costs, the elasticity of capital with respect to the user cost in this model is ¡1=´. Hence, for example, if´= 0:2, then ¡1=´= 5:0, so that a 10% increase in the user cost of capital implies a 50% decrease in the optimal capital stock.
To illustrate the e¤ect of changes in investor protection on the capital stock, we assume parameter values for´, ±, r, and ¡, respectively, of 0:2, 0:07, 0:10, and 0:0. The value of A is chosen to normalize K = 100 when investor protection is perfect (this corresponds to ® = 0 in equilibrium). Using equation ( In related research, Kumar, Rajan, and Zingales (2001) investigate the determinants of …rm size and …nd that their measure of judicial e¢ciency is an important explanatory variable. The numerical calculations in Table 6 are consistent with their evidence. Moreover, these results illustrate the endogeneity of the relationship between …rm size and inside ownership concentration proposed in section (4.3) as an explanation for the robust empirical relationship observed in Table 3. The calculations in Table 6 imply a relationship between ownership concentration and the log of capital is approximately linear with a slope of roughly ¡1:3, whereas the estimated coe¢cients in Table 3 range from ¡2:44 to ¡3:59.
Hence, the sign is correct and the values from this calibration exercise have the right order of magnitude. It is tempting to propose a model for …rm size by taking the log of equation (30) and, rearranging terms, regressing log K it on ownership concentration and other controls for other determinants of the cost of capital. Of course, ownership is endogenous, so it would not be appropriate to interpret ownership as a "determinant" of …rm size. Rather, this regression would simply recover the negative equilibrium relationship between …rm size and ownership.
Although the negative relationship between ownership and …rm size is a robust feature of the data, the problem with this proposed regression, unfortunately, is that "pro…tability" parameter, log A it , appears in the error term. This parameter is highly endogenous to both …rm size and ownership. Without instrumental variables to account for this omitted variable, such a regression provides biased estimates of the structural parameters. The speci…cation in equation (28), by contrast, do not su¤er from this bias.

Inside Ownership and Tobin's Q
The marginal value of capital, or marginal q, is the discounted marginal value (to the market) of an additional dollar of investment. That is, marginal q is de…ned as: Under zero adjustment costs, constant returns to scale, perfectly competitive product markets, and perfect investor protection, equation (6) says that marginal q equals one in equilibrium. It also follows immediately from the discussion of equation (17) that under imperfect investor protection and linear homogeneity of ¼, the equilibrium value of marginal q exceeds one. It is an easy algebraic exercise to extend this result to the case where the value function is homogenous of degree less than one, Ã < 1. 12 To map these statements about marginal q into statements about Tobin's average Q, we consider the general case where the one-period pro…t function ¼ is homogeneous of degree Ã < 1. Under this assumption, the relationship between marginal q and Tobin's Q is given 12 Homogeneity of degree Ã implies ¼ K = Ã ¼ K . This implication combined with equations (8) and (14) can be used to show q > 1.
In the special case that the pro…t function is linearly homogeneous (Ã = 1), we have the familar result that marginal q equals Tobin's Q. In the general case (Ã < 1), equation (32) implies Q it > q it . If Ã is a constant, the relationship between Q it and q it is linear. Thus, under fairly general conditions, weak investor protection implies the equilibrium value of Tobin's Q is greater than one. Moreover, by the relationship of marginal q to marginal pro…t and the logic of equation (21), the model predicts Tobin's Q is positively related to inside ownership concentration.
Our model thus provides an alternative explanation some of the results found with regressions of Tobin's Q on inside ownership (e.g, MÁrck, Shleifer, and Vishny, 1988;Mc-Connell and Servaes, 1990;Holderness, Kroszner, and Sheehan, 1999;and Himmelberg, Hubbard, and Palia, 1999, among others). A common interpretation for positive estimated coe¢cients on inside ownership in the above regression is "better incentives generate better performance." In this view, high values of Tobin's Q indicate "good performance," and therefore Tobin's Q should be higher for …rms with "good incentives," i.e., higher concentrations of inside ownership. McConnell and Servaes (1990), Himmelberg, Hubbard, and Palia (1999), and Demsetz and Villalonga (2001), however, raise various objections to this interpretation as well as to the practice of regressing Tobin's Q on ownership. The primary complaint is that ownership is endogenous. Our model addresses this problem by providing an empirical framework within which the consequence of this endogeneity can in principal, at least, be interpreted. In our model, high values of marginal q re ‡ect underinvestment resulting from low levels of investor protection, which in turn is positively correlated with ownership concentration. Hence equation (32) turns the traditional interpretation on its head; ownership concentration implies better incentives, but such incentives are necessary only when investor protection is weak. Ownership concentration and high values of Tobin's Q are merely joint symptoms of weak investor protection.
Because of the potential biases introduced by measurement problems with Tobin's Q (Himmelberg, Hubbard, and Palia, 1999;Demsetz and Villalonga, 2001), we do not attempt to estimate equation (32). For example, if the …rm's value function is not homogeneous of degree one (Ã < 1) due to, say, market power, then average Q does not equal marginal q. The discrepancy between the two stems from the fact that average Q values inframarginal rents on assets in place whereas marginal q concerns only the value of rents on the margin. This point holds for any other source of inframarginal rents, and applies to inframarginal costs as well. With …xed costs in production, for example, Tobin's Q can be less than marginal q. In short, average Q can easily re ‡ect substantial variation which is unrelated to marginal q. To make matters worse, if inframarginal rents are correlated with unobserved …rm-level or country-level investor protection variables, then the error term is correlated with the regressors, and in the absence of good instruments, least squares estimates of equation (32) are biased downward. By contrast, our adjusted sales-to-capital-based measure of marginal pro…t is robust to market power, …xed cost, and various other measurement issues that break the link between average and marginal q.

Financial Liberalizations
Our results suggest large potential gains from …nancial sector reforms that improve the level of investor protection. Most research on …nancial liberalization approaches the issue from an asset-pricing perspective which focuses on changes in the risk-free rate or the price of systematic risk (or both) as a consequence of improved international diversi…cation. As Shleifer and Wolfenzon (2000), among others, have pointed out, however, removing the barriers to capital ‡ows does not guarantee that capital ‡ows to its most e¢cient use unless international investors can be credibly convinced that investments will be repaid; the expected return to investors depends on the level of investor protection. For example, the estimates in Chari and Henry (2001) indicate that for a …rm operating in a market in which the covariance between the local and world market returns exceeds 0.01, …nancial liberalization causes a …rm-speci…c revaluation on the order of 3.4%. This revaluation occurs only for "investible …rms"; …rms which are "o¤ limits to foreign investors" bear no signi…cant relationship to di¤erences in local and world covariances.
The model in this paper formalizes this idea by providing quantitative guidance on the extent to which …rms are "o¤ limits" to investors. Recall from equation (12) that the …rst-order condition for capital is: If …nancial liberalization improves international diversi…cation, this implies a change in the stochastic properties of the market's SDF, M t+1 , which would change the risk-free rate, r f t+1 , and presumably lower the premium for systematic risk, h it . If investor protection were perfect, then we would have g it = 0 and h it = 1, in which case the only mechanism by which …nancial liberalization could a¤ect investment would be through changes in the risk-free rate and the repricing of systematic risk.
Under imperfect investor protection, however, the weight given to idiosyncratic risk is re ‡ected by the level of inside ownership. In the polar case for which investor protection is so weak that owners are autonomous (g it = 1, h it = 0), the e¤ects of …nancial liberalization on investment would have to operate indirectly through the e¤ects on the risk-free rate or the market's SDF caused by capital ‡ows out of the country. 14 More generally, equation (33) describes a range of intermediate cases for which the weights g it and h it fall somewhere between zero and one.
Equation (33) therefore provides an empirical framework for distinguishing the diversi…cation bene…ts from the investor protection reforms that often (to some extent) accompany …nancial market liberalizations. Intriguingly, and consistent with the evidence reported in Chari and Henry (2001), Bekaert, Harvey, and Lundblad (2001) …nd that the pre-existence of an Anglo-Saxon legal system magni…es the response of the investment-to-GDP ratio to …nancial liberalization events. This is precisely what the model presented in this paper predicts.

Conclusions
We investigate the cost of capital in a model in which investor protection determines the agency con ‡ict between inside managers and outside shareholders. Our principal empirical results con…rm two predictions of the model. First, the weaker is investor protection, the higher is the concentration of inside equity ownership. Second, the higher is the concentra-  Rajan, and Zingales (2001) have termed the "technological" and "organizational" theories of the …rm. We provide new evidence consistent with the view that "organizational" factors (like investor protection) are important determinants of …rm size. Third, we have formally argued that because weak investor protection leads to underinvestment, the marginal pro…t of capital is not driven down to its …rst-best level and therefore Tobin's Q is greater than one in equilibrium. In addition, because inside ownership concentration is higher under weak investor protection, the equilibrium relationship between inside ownership and Tobin's Q is positive. Subject to quali…cations regarding possible discrepancies between average and marginal Q, these results provide a new interpretation for ownership-performance correlations which di¤ers from previous explanations. Fourth, our model helps to shed light on the real economic e¤ects of …nancial liberalizations. In particular, it helps to formalize the widely recognized fact that while lowering international barriers to capital ‡ows is obviously a necessary condition for liberalization, it is not su¢cient; capital will not ‡ow unless adequate investor protections are in place. Existing empirical work already provides evidence for this intuition which is succinctly captured by our expression for the cost of capital. The ratio of the book value of debt to the book value of assets The ratio of the book value of debt to the market value of assets The fraction of equity held by insiders (in Worldscope, the variable "closely held shares") Table 2 The log of firm sales, where sales is measured in constant U.S. dollars

Percentiles
The ratio of R&D expenditures to sales A dummy variable equal to one if R&D is missing, zero otherwise

Determinants of Inside Ownership Concentration
Coefficients from regressions of inside ownership on country-level and firm-level measures of investor protection. Constant terms are not reported. Standard errors (in parentheses) adjust for heteroscedasticity and within-firm serial correlation. Statistical significance levels are denoted by stars, where ** and * denote significance at the one and five percent levels, respectively (two-tailed tests).  Table 4 IV OLS

Estimates of the First-Order Condition for the Capital Stock
Coefficients from regressions of the marginal return on capital (MPK) on inside ownership (equation (11) in the paper). Constant terms and dummy variables are not reported. Column (2) repeats columns (1) adding country-specific time dummies, and column (3) repeats column (2) adding industry dummies. Column (4) repeats column (3) using the sample of the 150 largest firms in each country. Column (5) repeats column (3) using three lags of all variables as instruments. Column (6) adds the log of sales to column (5). Standard errors (in parentheses) adjust for heteroscedasticity and withinfirm serial correlation. Statistical significance levels are denoted by stars, where ***, ** and * denote significance at the one-, five-and ten-percent levels, respectively (two-tailed tests).

Estimates of Euler Equations and Leverage Effects
(1) (2) (3) (4) (5)  Model extensions to the regression of the marginal return on capital (MPK) on inside ownership (equation (11) in the paper). Column (1) reproduces model (5) from Table 4 for comparison. Columns (2) and (3) add leverage, measured as the ratio of the book value of total liabilities to total liabilities plus equity (using the market and book values of equity, respectively). Columns (4)-(6) report Euler equation estimates to capture dynamics in MPK resulting from adjustment costs. All specifications are estimated with industry-and country-specific year dummies, and all use three lags of the dependent and explanatory variables as instrumental variables. Standard errors (in parentheses) adjust for heteroscedasticity and within-firm serial correlation. Statistical significance levels are denoted by stars, where ***, ** and * denote significance at the one-, five-and ten-percent levels, respectively (two-tailed tests).  Table 6 Equilibrium magnitude of underinvestment implied by observed ownership concentration under alternative values of the idiosyncratic risk premium, γ-Γ Solutions for π K and K assuming η=0.2, r+δ+Γ=0.18, and values for α and γ-Γ as indicated in the respective row and column headings. The value of the "profitability" parameter A is chosen to normalize K equal to 100 in the benchmark case in which perfect investor protection (i.e., α=0).