Getting Deals Done: The Use of Social Networks in Bank Decision-Making

by Mark S. Mizruchi, Linda Brewster Stearns
Getting Deals Done: The Use of Social Networks in Bank Decision-Making
Mark S. Mizruchi, Linda Brewster Stearns
American Sociological Review
Start Page: 
End Page: 
Select license: 
Select License


University of Michigan University of California, Riverside

Economic actors confront various forms of uncertainty making decisions, and how they deal with these obstacles may affect their success in accomplishing their goals. This study examines the means by which relationship managers in a major commer- cial bank attempt to close transactions with their corporate customers. It is hypoth- esized that under conditions of high uncertainty, bankers will rely on colleagues with whom they are strongly tied for advice on and support of their deals. Drawing on recent network theov, it is also hypothesized that transactions in which bankers use relatively sparse approval networks are more likely to successfully close than are transactions involving dense approval networks. Both hypotheses are supported. Bankers are faced with a strategic paradox: Their tendency to rely on those they trust in dealing with uncertainty creates conditions that render deals less likely to be closed successfully. This paradox represents an example of unanticipated conse-

quences of purposive social action.

APRIMARY GOAL of sociologists has been to demonstrate that the structure of social relations (networks) has consequences for behavior and outcomes. Re- cently, researchers have described actors' personal networks in terms of social capital, which they see as a resource that helps ac- tors achieve particular ends. Increasingly, social capital is viewed not only as a re- source possessed by virtue of our position in

Direct all correspondence to Mark Mizruchi, Department of Sociology, University of Michigan, Ann Arbor, MI, 48 109- 1382 (, or to Linda Brewster Stearns, Department of Sociology, University of California, Riverside, Riverside, CA 92521-0419 ( Research was supported by the Citigroup Behavioral Sciences Research Council and the Russell Sage Foundation. We thank Wayne Baker, Ron Burt, Daniel Byrd, William H. Greene, Mark Granovetter, Shin-Kap Han, Don Palmer, Yu Xie, the ASR Editors and reviewers, and colloquium audiences at Carnegie Mellon University, Cornell University, Harvard University, the University of Michigan, North- western University, and the 2001 Strategic Man- agement Conference at Stanford University for their helpful comments and suggestions.

a social structure, but also as something that, within limits, we can create, or tailor, to serve our goals. But even if we construct our own social networks, the structures that we create may not have the consequences that we anticipated. In some cases, what appears to be a perfectly reasonable use of social capital has consequences that are the oppo- site of those we intended.

We examine the means by which organi- zational actors-members of a major com- mercial bank-use their social networks within the bank to close deals with their corporate customers. Our analysis consists of two stages. In the first stage, we propose hypotheses to account for the types of net- works bankers construct. Building on a siz- able literature in organizational sociology, we suggest that high levels of uncertainty will lead bankers to rely on their social ties within the bank. In the second stage, we ar- gue that the structures of the networks cre- ated by these social relations affect the probability that a banker will successfully close a transaction.

The simultaneous focus on the causes and consequences of networks has been an on- going concern in organizational research (Gulati 1998; Powell, Koput, and Smith- Doerr 1996), and increasingly authors ac- knowledge that networks can have negative as well as positive outcomes (Baker and Faulkner 1993; Gargiulo and Benassi 2000; Labianca, Brass, and Gray 1998; Uzzi 1996). Most researchers who have focused on the initiation of network ties, however, have as- sumed that actors create networks to solve problems, such as filling a need for skills or resources. The successful construction of such networks is then assumed to have posi- tive, anticipated consequences for the actor. This view is well captured by Lin (2000), who defines social capital as "investment and use of embedded resources in social re- lations for expected returns" (p. 786). A con- siderable body of research demonstrates the value of the strategic use of social networks (for a review, see Burt 2000). There are times, however, in which the returns from social ties are unexpected and not antici- pated by those who created them. Ties de- veloped with the goal of achieving one out- come may have the opposite effect. We show that the nature of the networks that bankers construct to solve a set of problems can ac- tually reduce the bankers' probability of be- ing successful. This suggests that the process of network formation may yield unexpected as well as expected returns.

We focus on two issues: First, the relation between the level of uncertainty and the use of social networks within the bank; second, the extent to which the nature of bankers' social networks affects the outcome of a transaction.

The concept of uncertainty, which March (1994) defines as "imprecision in estimates of future consequences conditional on present actions" (p. 178), has played a prominent role in organizational theory. From Simon's (1947) classic work on bounded rationality to a subsequent range of very different formulations (DiMaggio and Powell 1983; Meyer and Rowan 1977; Pfeffer and Salancik 1978; Williamson 1975), uncertainty reduction and uncertainty management have been at the center of dis- cussions of inter- and intrafirm relations.

Some organizations have as their primary basis of existence the management of uncer- tainty. A good example of such an organiza- tion is a bank. Banking thus represents an especially good site for examining the way actors deal with uncertainty.

Uncertainty can have a number of dimen- sions. A deal may be risky because a cus- tomer is in a precarious financial situation or has recently experienced poor financial performance. The banker may have varying degrees of trust in his or her customer. The transaction itself may be highly complex, in- corporating a large number of unknowns. Bankers deal with these uncertainties on a daily basis, and one way they manage them is by consulting their colleagues. A consid- erable literature suggests that social ties like these can mitigate the uncertainties both among and within organizations. Long-term social relations between a firm's sales agents and purchasing agents and those of customer and supplier firms can reduce uncertainties and thus the transaction costs associated with interfirm business (Granovetter 1985). Mechanisms such as director interlocks can reduce uncertainties associated with resource dependencies (Pfeffer and Salancik 1978). The creation of specific organiza- tional units, such as investor relations of- fices, can signal to the larger environment that an organization is a responsible actor, thus increasing its legitimacy and claim to societal resources (Meyer and Rowan 1977). And social relations among firms' officers can reduce the uncertainty associated with a range of organizational innovations and be- haviors, from merger financing (Stearns and Allan 1996) and takeover defense strategies (Davis 1991) to contributions to nonprofit organizations (Galaskiewicz 1985) and po- litical candidates (Mizruchi 1992).

The networks that actors create to cope with uncertainties may have impacts of their own, however. A growing literature, stretch- ing back to Granovetter's (1973) classic es- say, suggests that weak ties provide a wider range of information than do strong ties. Burt (1992) suggests that actors whose so- cial contacts are disconnected from one an- other have higher success rates than do ac- tors with dense personal networks. Strong ties tend to be densely connected. Because actors may turn to those they trust under conditions of high uncertainty, this normal response to uncertainty may lead to deleteri- ous consequences. If strong ties lead to dense networks with a limited range of in- formation, and if weak ties lead to sparse networks with a broader range of informa- tion, then the counterintuitive decision to rely on weak ties under conditions of high uncertainty may actually yield superior out- comes.

The site for our study is the corporate bank- ing unit of a leading multinational commer- cial bank, which we refer to as UniBank. The bankers in this unit deal with the bank's approximately 1,400 corporate customers, most of whom are large, multinational firms. In preparation for our study, we gathered in- formation on UniBank policies by canvass- ing annual reports over the past several years and examining internal documents. We also had extensive discussions with two contact officers within the bank, and conducted 14 in-depth, open-ended interviews with bank officers in three U.S. locations (Chicago, Los Angeles, and New York). Interviews ranged from one to four hours each. These discussions familiarized us with the bank's deal-making processes.

Bankers pursue several forms of informa- tion when processing a transaction. First, they seek out the most complete and high- quality information available on the cus- tomer's present and future financial condi- tion, including formal credit information, both external (Standard and Poor's reports and ratings) and internal (the bank's own credit rating model). Second, bankers draw on information obtained through their own and colleagues' experiences with the cus- tomer, including their knowledge and trust of the customer firm's management.

The bankers in our study are sales people. They sell by responding to clients' requests as well as by creating demand among these clients for the bank's many products. These products can be divided into four classifica- tions (examples in parentheses): lending (lines of credit and project finance), trading (derivatives and currency exchange), capital market services (syndication and securiti- zation), and transactional services (cash management and custody). The bankers' goal is to close deals. They are constrained, however, in the terms that they can offer their customers. Before a transaction at UniBank can close, the deal normally requires the approval of at least three officers within the bank. This requirement is the bank's formal mechanism for ensuring that the bankers sell to customers only those deals that meet the bank's targeted rate of return. This approval is no guarantee that the bank will achieve the estimated rate of re- turn, only that the deal is expected to do so. Although the banker and the officers from whom he or she seeks approval are con- cerned with closing the deal and its rate of return, the banker's primary concern during the approval process is with closure, while the officers' primary concern is with the re- turn rate. Once a transaction has reached ap- proval status, their goals converge: Both want the deal to close. It is important, there- fore, to distinguish between a banker "suc- cessfully closing a deal," that is, obtaining business for the bank, and a "successful deal," that is, a deal that upon completion provides the bank with the expected rate of return. Our concern is with closing the deal.

To secure internal approval requires that the client be assigned a credit rating. The bank calculates credit ratings by combining Standard and Poor bond-rating data with its own information. During the approval pro- cess, the transaction is evaluated using a two-dimensional table known as the ap- proval grid. The grid consists of cells deter- mined by a combination of the customer's credit rating and the bank's total and mar- ginal (based on the size of the current deal) exposure to the customer. This grid deter- mines how high in the bank's hierarchy the banker must go to secure approval for a deal.

As noted above, each credit transaction re- quires the signatures of a minimum of three officers, at least one of whom must be a se- nior credit officer at the level designated by the grid. These senior credit officers, of whom there are approximately 500, are re- quired to have 10 years of banking experi- ence; at least two of them must be at Uni- Bank. There are three types of senior credit officers-senior bankers, risk managers, and executive vice presidents-spanning four

levels. Above the senior credit officers is a six-member credit policy committee. Above this committee are "contact executives," key senior executives such as the vice chairman of the bank.

Bankers involved in transactions use their social networks within the bank to seek ad- vice from peers and superiors who have knowledge based on product expertise or past experiences with the customer. In the context of banker decision-making, it is use- ful to think in terms of two types of net- works. On one hand, there are what we call information networks. These are social rela- tions that bankers use to secure knowledge about the status of particular firms or prod- ucts, or appropriate ways to structure a trans- action. On the other hand, there are what we call approval networks. These are social re- lations that bankers use to gain both confir- mation and support for the transaction.

An example of the use of information net- works appeared in one of our interviews with a senior banker. The banker recounted to us a hypothetical case involving a major customer. In a typical case, the banker brings in a product specialist in the particular area, such as syndications, cash management, or foreign exchange. The teams vary based on the company involved and the type of trans- action requested. "It could be a straight deal such as cash management or a never-before- done deal," the banker told us. Often it is a competitive situation with other banks, in which the bank must "make its deal appear 'prettier' than the others'." In these situa- tions, the banker will consult colleagues within the bank, often those with whom he or she has consulted previously. As he noted, "That's where we get a lot of the network going." The banker will ask, "Gee, have you seen something like this before. . . ? You talk to some people. One of the reasons they come to me is the exact same thing. They want to see what I think, because, maybe I've been doing it twice as long as they have." This example, typical of many we heard, suggests that networks are often used to reduce uncertainty among bankers.

An example of the use of approval net- works was recounted to us by a risk man- ager. A banker was putting together an ac- quisition deal involving a major manufactur- ing firm. Because the company had a poor

1 credit rating, the deal required approval by a

, Level 1 credit officer (the bank's term for its

, highest level) plus the credit policy commit-

I tee and a contact executive. Initially, the ap- propriate Level 1 signator was "uncomfort- able" with the deal. The banker and risk manager then assembled a meeting with the Level 1 officer and the contact executive. The latter said he was "OK with the deal" (that is, he thought the deal was within ac- ceptable risk), but he would not sign it un- less the Level 1 officer also signed. The banker and risk manager then approached a credit policy committee representative. They "got him comfortable with the deal" as well, but as with the contact executive, the credit policy person agreed to sign only if the Level 1 officer signed. Eventually, the Level 1 officer signed, thus securing the deal. In cases like this, network ties often determine the specific persons whose support is solic- ited. In trying to get the required signatures, a senior credit officer told us, he seeks out those "higher-ups [he believes] will be fa- vorable to the deal." Once bankers decide that a particular deal should go forward, they attempt to manage the disagreements sur- rounding the deal by using approval net- works to obtain the necessary support. How much discretion do the bankers have in constructing their networks? Clearly, they, like all actors, are constrained to some ex- tent: One cannot develop network ties with actors who are unavailable. The formal orga- nizational structure as well as the nature of the particular deal also play a role. Most deals require bankers to work with product specialists. There is, however, no formal means for assigning a product specialist to a deal. Although bankers are assigned to a customer's account, product specialists are not. Rather, they are assigned to a product specialty area, such as lending or trading. The banker therefore has some discretion in selecting the product specialists with whom he or she will work on a given deal. The more complex the deal, the more likely the banker will consult several product specialists. In constructing their approval networks, most bankers are expected to seek support from their division head and, in many cases, the group's risk manager. Although deals gener- ally require three signatures, most deals have more than three. There are few limitations

regarding whom can be consulted for infor- mation or support, although a banker's rank may affect his or her access to some indi- viduals. One illustration of how bankers' net- works can vary across deals is the fact that our results are virtually identical with or without controls for within-banker variance. This suggests that there is as much variance among deals by individuals as there is across all bankers. Our bankers may not make their networks exactly as they please, but they clearly do make them.

Our analysis proceeds in two stages. First, we examine the extent to which uncertainty affects bankers' use of social networks. Sec- ond, we examine the effects of these net- works on whether a banker successfully closes a deal with a corporate customer. In the first stage, we suggest that high levels of uncertainty will lead bankers to rely on their ties within the bank. We examine as endogenous variables the strength of the ties that bankers use in both their information and approval networks. In the second stage, we use as our dependent variable a dichoto- mous measure indicating whether a given deal was closed. Both our exogenous and in- tervening variables are expected to affect the outcome measure (closing a deal) in various ways.

Our primary exogenous variable is the level of uncertainty, which we divide into two components: economic uncertainty and lack of trust. Economic uncertainty is defined as the banker's perception of the degree of risk to the bank's capital. Lack of trust is de- fined as the extent to which the banker has doubts about the individuals in the customer firm with whom he or she interacts

We have suggested that in cases where un- certainty is high, bankers will consult with colleagues within the bank to gain more in- formation about the customer or the type of transaction; that is, they will use their infor- mation networks. In situations in which a banker has not yet determined whether to support a deal, his or her colleagues may provide negative as well as positive informa- tion about the company or the deal. The banker's primary concern in using informa- tion networks is thus to assemble additional information, not necessarily to gain support for a decision in which he or she already has a stake. It is in situations in which the banker lacks knowledge of the customer or requires advice on the structuring of a deal that he or she will be more likely to invoke internal in- formation networks. In fact, most of the more senior bankers had long-term relations with their customers. For these bankers, the use of information networks typically in- volved questions about the deal rather than the customer.

Which colleagues will a banker consult? On one hand, it is known from network theory (Granovetter 1973) and evidence (Granovetter 1974) that actors gain more in- formation from those with whom they are relatively weakly tied. On the other hand, it is not clear that actors are consciously aware of the benefits of weak ties, and there is also evidence that under uncertain conditions ac- tors rely most heavily on those they trust (Kanter 1977). The data from our prelimi- nary interviews are consistent with this lat- ter suggestion. To the extent that the benefits of weak ties are unknown and the value of strong ties is apparent, we believe that, when faced with a difficult situation, actors will be more likely to turn to those with whom they are strongly tied.

Hypothesis 1: The higher the uncertainty in a transaction, the stronger the ties be- tween the banker and those he or she consults for information.

High levels of uncertainty also increase the probability of ambiguous interpretations (March 1994: 175-219). In highly uncertain environments, the number of potential vari- ables affecting an outcome increases. Thus, the possibility for multiple interpretations increases for decisions involving high uncer- tainty. Once bankers have made a decision to pursue a deal, they actively seek support

1 for their position. This is done through the

' use of approval networks. We expect that bankers will seek out approval networks in much the same way that they seek out infor- mation networks. This means that under con- ditions of uncertainty, bankers will seek ap- proval from those they trust.

1 Hypothesis 2: The higher the uncertainty in a transaction, the stronger the ties be-

tween the banker and those from whom he or she seeks approval for a deal.

A third possible source of uncertainty is the complexity of the deal. Complex deals, which are distinguishable from risky ones, necessarily require consultation with a rela- tively large number of colleagues. Even if bankers might prefer to rely for advice on those they trust, the complexity of the deal may lead them to consult a larger network, which may involve a weaker set of ties. We therefore do not hypothesize a specific ef- fect for deal complexity on the strength of a banker's ties. We do, however, include this variable as a control.

The dependent variable in the second stage of the model is whether a transaction closed. This was the primary goal of all of the bank- ers with whom we spoke. It is important to distinguish closure from approval. Situations in which bankers took a deal through the for- mal process but ultimately failed to gain in- house approval were rare in our study; we were able to document only two such cases. Our interviews suggested that self-censoring based on preliminary "testing of the waters" was the primary reason for this. Far more common are situations in which the banker gets his or her deal approved within the bank but fails to close on it. This can occur for a number of reasons. The customer may decide that the terms of the deal are unacceptable and then choose to work with a bank that of- fers more favorable terms. In some cases, the bank and customer simply abandon the deal and consider an alternative approach. In other cases, a customer's situation changes in the midst of negotiations. The firm may be acquired by another or engage in an acquisi- tion of its own. The primary determinant of whether a deal successfully closes, however, is the banker's ability to construct an agree- ment with which his or her officers are satis- fied and that is attractive to the customer.

Of what does such an attractive deal con- sist? Consider an example in which a cus- tomer requests a loan for an amount or inter- est rate that the bank finds unacceptable. Rather than reject the request, the bank's policy is to restructure the deal to meet the customer's needs while also generating an acceptable rate of return for the bank. In such a case, the customer may have an over- seas facility in which it wants to invest. The banker, working with members of his or her approval network, may find that by restruc- turing the deal to draw the funds from the bank's overseas branch, the customer can benefit from currency and tax advantages. The customer receives the amount and price it wants, and the bank receives an acceptable interest rate and a larger up-front fee. There is no guarantee that restructuring a deal will produce a mutually beneficial outcome, but this is the goal of the banker and the bank. Our concern is with the factors that predict a banker's ability to work out a mutually ac- ceptable arrangement for a given deal.

Our argument suggests that to the extent that economic uncertainty and trust affect the closure of a deal, they would do so through their effects on the use of informa- tion and approval networks. At the same time, it is possible that uncertainty has a di- rect effect on closure. We expect, for two reasons, that deals with higher levels of eco- nomic uncertainty will be less likely to close. First, bankers may abandon a high- risk deal early in the process because they believe they will experience difficulty gain- ing approval for the deal. Second, when bankers do pursue such deals, there is a greater chance that the deal will be restruc- tured during the approval process in a way that makes it more difficult to sell to the cus- tomer, as, for example, in charging a higher interest rate or imposing additional restric- tive covenants. We also expect that bankers will have a more difficult time closing deals with customers with whom they have less than complete trust. This suggests that the banker's trust of the customer will be posi- tively associated with closure. And although we did not hypothesize a specific effect of complexity on the strength of a banker's ties, it is likely that bankers will have more diffi- culty closing highly complex deals, suggest- ing that we should observe a negative effect of deal complexity on closure. This discus- sion suggests the following hypothesis:

Hypothesis 3: The higher the uncertainty in

a transaction, the less likely the transac-

tion will close.

The effect of information networks on the closure of a deal is not necessarily straight- forward. To the extent that these networks are

large and nonredundant, the banker will have a wider range of data on which to both struc- ture and evaluate a deal. Network size refers to the number of individuals consulted by a banker. Redundancy refers to the density of the actor's personal network-the extent to which those consulted by the actor are tied to one another. When personal networks are dense, actors are likely to receive the same or similar information, because this informa- tion circulates among the same group of people. When personal networks are sparse, actors are likely to receive a greater range of information, because members of the net- work tend to be tied to a diverse set of alters (Burt 1992; Granovetter 1973). To the extent that more information from a range of sources is better than less information, sparse networks should benefit the course of a deal on which the banker is working. In deciding whether to pursue a deal, a different outcome is possible, however, because members of one's network may provide negative as well as positive information about the deal. If this is the case, then there would likely be no sys- tematic association between the use and na- ture of information networks and the prob- ability that a deal will be closed. Because of the potential crosscutting influences of mem- bers of one's information network, it is pos- sible that a broad, sparse information net- work will increase the probability of closure, but it is also possible that there will be no association between the two variables. We therefore do not have a specific hypothesis about the relation between the density of a banker's information network and the prob- ability of closure, although we include this variable in our model.

Approval networks, on the other hand, are used when a banker wants to gain confirma- tion for a deal that he or she already sup- ports. As with information networks, ap- proval networks vary in both size and den- sity. Whether a banker is able to secure sup- port for a deal may depend on the range and breadth of the colleagues he or she enlists. Sparsely connected groups are likely to con- tain a wider range of views and expertise than are more densely connected groups. This means that ideas supported by members of low-density groups will tend to have re- ceived more criticism and questioning and will benefit from a greater range of insights.

To the extent that a banker is able to gain support from a broad range of colleagues, he or she will be able to present a "better" deal-one that is attractive to both the cus- tomer and the bank.

Hypothesis 4: The lower the density of the approval network consulted by a banker, the more likely the transaction will close.

Finally, related to the question of network redundancy is the extent to which an actor depends on a single individual. Burt (1992) argues that actors whose bases of informa- tion and social support are restricted by a strategically located individual are likely to be disadvantaged in the acquisition of orga- nizational resources. A banker who is depen- dent on a single person for closing a deal may have few others to whom he or she can turn should this person fail to provide sup- port. The extent to which an actor is highly dependent on a single individual, which Burt refers to as the degree of hierarchy, can be measured by examining the level of inequal- ity in the actor's network. Hierarchical net- works are dominated by one person or a small number of persons. Nonhierarchical networks tend to give actors more alterna- tive sources of support. Burt (1992) shows that corporate managers embedded in hier- archical networks had longer times to pro- motion than did those whose networks were less hierarchical. This suggests that bankers whose approval networks are hierarchical will, on average, have a more difficult time gaining support for, and closing, their deals.

H-ypothesis 5: The less hierarchical the ap- proval network consulted by a banker, the more likely the transaction will close.


Our units of analysis are specific transac- tions. These transactions may include any one, or combination of, the products mentioned above, including lending, trading (de- rivatives and currency exchange), capital market services (syndication and securiti- zation), and transactional services (cash management and custody). From May 1997

through March 1999, we conducted semi- structured interviews with 91 of the 110 bankers in the bank's "global relationship banking" unit. This unit is responsible for handling the approximately 1,400 multina- tional corporations that the bank had tar- geted as its corporate customers. The bank- ers with whom we spoke represented 16 business units in two domestic locations.

Our goal was to interview each banker twice. We were able to successfully re-inter- view 82 of our original 91 respondents. Most of the others were lost due to attrition (such as moving to another location in the bank, leaving the bank, or taking maternity leave). At the initial interview, we asked each banker to describe three transactions in which he or she was currently involved. Bankers were asked to provide us with a range of deals, from simple to complex, and encompassing a range of the kinds of activities in which they typically engage. We then asked bank- ers to provide specific information about the deals and their relations with the companies. At the follow-up interview, we learned the outcome of the deals and asked a number of questions, including the name generators that yielded the network variables2 All of the ini- tial interviews and a majority of the follow- up interviews were conducted in person. Be- cause the outcome of deals was not always known at the first follow-up, we conducted some of the follow-up interviews by phone. We secured at least some information on 230 deals at their initial stage. We collected out- come information on 194 deals. The remain-

The business units consist of two regional of- fices (New York and Chicago) and 14 industry units. The industry units are automobiles, avia- tion, banks, branded consumer, chemicals/phar- maceuticals, communications, electronics, global energy, global power, insurance, investment bankslmanaged funds, retail, shipping, and tech- nology.

The bank maintained a "deal pipeline," a file in which bankers were expected to register all deals on which they were currently working. Our original goal was to use the deal pipeline to gen- erate the deals for our study, which would have allowed us to collect a random sample of deals. In the course of our preliminary interviews, how- ever, we learned that the deal pipeline was an in- valid indicator of the banker's current portfolio. Because there was no apparent sanction for

der were lost due to attrition. Based on their primary component, 27.3 percent involved loans to be held by the bank, 22.7 percent involved capital market services, 16.5 per- cent involved transactional services, 16.0 percent involved securitized loans (to be sold off by the bank), and 14.4 percent involved trading. The remainder involved a combina- tion of products. Of these 194 deals, 96, or slightly under 50 percent, were successfully closed. Loans were slightly more likely than the other products to close, but the difference was not statistically significant. Missing data on particular variables reduced the number of usable observations to 173 or fewer for our analyses.

Our primary exogenous variables are our measures of uncertainty, which include eco- nomic uncertainty, lack of trust in the cus- tomer, and the complexity of the deal. Economic uncertainty is defined by the banker's perception of the financial risk of a deal. One possible indicator of risk is the approval grid, which views risk in terms of a combination of the bank's total exposure to the customer, the size of the deal, and the customer's credit rating. The bank uses a 1 to 6+ rating scale (with 6+ the highest risk), which roughly corresponds to Standard and Poor's ratings. There are two reasons that this rating is not a suitable measure for our purposes. First, the grid rating is normally assigned at a rela- tively late stage in the approval process. Be- cause of its role in our model as a predictor of the colleagues the banker consults, we wanted to gauge the banker's perception of risk in the early stages of the deal. This was especially important because in a significant number of cases, the approval level did not follow the formal scheme but varied as a re- sult of negotiation by the banker and his or her colleagues. This meant that the actual approval level may have been as much a con- sequence of the banker's approval network

underreporting, most bankers neglected to regis- ter many of the deals in which they were in- volved, because of time pressures as well as con- cern for their success rate. The best alternative was to ask the bankers to generate the deals in a way designed to ensure sufficient variation. We do not claim that this produced a random sample of deals, but we are confident that it provided a wide representation of the kinds of services that the bankers provide.

as it was an indicator of risk. Second, al- though the bankers were aware of both the customer's credit rating and the bank's expo- sure in the particular deal, they were often unaware of the bank's current total exposure to the customer, especially if the bank was involved in other ongoing deals with the cus- tomer. Our goal was to identify a measure of risk that was specific to the particular deal and that reflected the banker's perception prior to reaching the approval stage. On the basis of its similarity to the approval grid, we operationalized economic uncertainty as the product of the bank's exposure in the deal and the customer's credit rating, converted to logarithms (base e) to adjust for skew- new3

Lack of trust is defined as the extent to which the banker has doubts about the indi- viduals in the customer firm with whom he or she interacts. We asked bankers to tell us, on a 4-point scale (4 is high), the extent to which they "trust the individuals at the cus- tomer firm to do what they say they're going to do." The bankers' trust of their customers tended to be high: Nearly 50 percent of all trust relations were given the highest score of 4, and more than 80percent of the remain- der received scores of 3. We determined on the basis of this finding that the most impor- tant distinction was between a score of 4 and all others. We therefore combined scores of 1 through 3 into a single category and treated customer trust as a dichotomous variable, with 1 indicating high trust (no doubt) and 0 indicating low trust (at least some doubt).

Complexity of the deal was determined by asking bankers to rate the deal, relative to those on which they generally work, on a

An analysis of the approval grid indicated that the level of exposure in the deal and the bank's total exposure to the customer at given approval levels were correlated .99. In other words, in using the grid to predict the required approval level, the deal-specific exposure and to- tal exposure were interchangeable. We operation- alized risk as the product of the customer's risk rating and the bank's exposure in the particular deal because the outcome cells of the grid re- sembled a product of the customer's risk rating and exposure. This indicator had a .92 correlation with the approval level on the grid, indicat- ing that it was a good approximation of the bank's conception of risk.

scale of 1 to 5, with 1 being very simple and routine and 5 being highly complex.

The network variables were generated by asking respondents to provide the first names or initials of up to eight individuals whom they consulted for information (infor- mation networks) or support (approval net- works). Each banker had separate informa- tion and approval networks for each deal. We asked bankers to rate the strength of each relation, both between them and each alter and among the alters themselves, on a 1 to 4 scale: 1 = infrequent work colleague, 2 = moderately frequent work colleague, 3 = frequent work colleague, and 4 = frequent work colleague who is also a personal friend (de- fined in terms of knowing one another's family or having entertained in one another's homes). There were no cases in which bank- ers mentioned personal friends who were not also frequent work contacts. Zeros were oc- casionally used by bankers to identify cases in which alters had no contact or had no knowledge of the other's existence. We com- puted three separate measures based on this scale. The first, the strength of the banker's ego network, was computed for the banker's direct relations with members of his or her information and approval networks. It was computed by the formula

where E, equals the strength of actor i's ego network, S,, equals the strength (on the 1 to 4 scale) of the actor's relations with each al- ter j, and N equals the number of alters. The sum of the values for each of the banker's direct relations is divided by the number of ties times four, since 4N is the highest pos- sible sum.

The second measure, which we refer to as alter network density or (simply "density"),

I is the level of connectivity among those in the banker's network for the particular deal.

, It consists of the weighted strength of rela- tions among those with whom the banker is tied. This is computed by the formula

I where Di equals the density of banker i's net- work, Sjk equals the strength (on the 1 to 4

scale) of each of actor i's direct ties j with i's other ties (k), and 2(I@ -N) equals the number of possible ties among banker i's di- rect ties [(fl-N)/2] multiplied by 4, since 4 is the maximum strength of a given rela- tion. The first of these measures, strength of the banker's ego network, was used to test Hypothesis 1 (for the information network) and Hypothesis 2 (for the approval network). The second measure, the density of the banker's network, was used to test Hypoth- esis 4, as well as to measure the density of the bankers' information network^.^

The use of different network measures as dependent variables in Hypotheses 1 and 2 from those used as independent variables in the analysis of closure is consistent with ex- isting theory as well as the model we de- velop. As Burt (1992, chap. 1) suggests, however, strong tie networks are not neces- sarily dense nor are weak tie networks nec- essarily sparse. Still, actors who seek out re- lations with strongly tied alters will tend to have denser alter networks than will those who seek out relations with more weakly tied alters. The correlations in our data be- tween information network direct tie strength and density were .47 for all deals and .58 for the deals that reached the ap- proval process. The correlation between ap- proval network direct tie strength and den- sity was .70. Although we doubt that they were consciously attempting to construct specific types of alter networks, there were occasions in which bankers approached col- leagues whose support was necessary to gain the support of other important colleagues. To ensure the tenability of our two-stage model, we computed additional equations that treat alter network density in the information and approval networks as endogenous variables in the first stage of our analysis (Hypotheses 1 and 2), and ego network information and approval network strength as exogenous

This measure is equivalent to the standard measure of network density (the number of ac- tual ties divided by the number of possible ties) weighted by the strength of tie. The conventional measure assumes a binary network, in which re- lations are either present or absent. In our net- works, the overwhelming majority of actors had at least some connection. The primary source of variation is the relative frequency of their inter- action.

variables in the second stage (Hypothesis 4). Although the two variables are highly corre- lated, we also inserted ego network strength into some of our equations simultaneously with alter network density.

Our third network measure, the hierarchy of a banker 's network, was operationalized as the coefficient of variation of the strength of the banker's direct ties. This is slightly different from the definition used by Burt (1992), but the coefficient of variation was identified by Allison (1978:877) as a valid measure of inequality for variables of this type. Although we did not develop a specific hypothesis about the effect of the hierarchy of the information network, we included this variable as a control.

We also controlled for three variables that might affect the outcome of a deal indepen- dent of our hypothesized factors: the banker's rank, whether UniBank was the pri- mary bank (or lead bank) for the customer firm, and the respondent-reported degree of consensus within the bank on the deal.

The bankers in our study occupied five different levels within the firm, from assis- tant vice president to managing director. A banker's rank within the bank can affect not only the kinds of deals on which he or she works, but also the kinds of networks he or she is able to construct. One lower-ranked banker, for example, complained to us about the difficulty of having his phone calls to colleagues returned, presumably because of his low status. There was little indication from our interviews that highly ranked bank- ers were either more or less likely to work on deals that had a high probability of suc- cess. It is appropriate to include this variable as a control, h~wever.~

In addition to the banker's rank, we exam- ined two person-specific variables, gender (29 percent of our deals were handled by women) and the banker's length of service dealing with the customer. The latter is a useful proxy for length of service with the bank, but has the advantage of being deal-specific. Gender was not signifi- cantly associated with either the strength of in- formation or approval network ties or deal out- come in any of the equations, nor did its pres- ence in the model affect the significance of the other coefficients. Length of service with the cus- tomer had a significant effect on the strength of information and approval network ties, but its in-

Similarly, whether UniBank is the customer's primary bank may affect whether the bank gains the customer's business on a particular deal. It is therefore important to control for this variable. Our indicator for lead bank was a three-level ordinal variable, with 0 indicating that UniBank was not the customer's lead bank, 1 indicating that UniBank was one of the customer's lead banks, and 2 indicating that UniBank was the customer's sole lead bank.

The consensus variable, which we mea- sured on a l to 5 scale (with 5 indicating unanimity), was designed to examine the ex- tent to which the banker's peers were "on the same wavelength" regarding the deal.

Finally, because the definition of network density includes -N in its denominator, density is in part a function of size-large networks tend to be relatively sparse. It is therefore possible that any effect we find for density is simply a function of the number of alters in a banker's network. Even if this were the case, our interpretation of the net- work effect could be similar, in that the larger the number of bankers from whom the banker gains support for a deal, the greater the likelihood that he or she received a broad range of feedback. But a large approval net- work could also be viewed as an additional indicator of risk, in that superiors may be more willing to support a high-risk deal when a substantial number of others have al- ready signed on (as in our example from our interview referred to above). Considering size and density simultaneously therefore al- lows us to examine the effect of the network structure independent of the previously un- explained factors that lead to large networks. We therefore include the size of the banker's networks as controls, where size is simply the number of alters named by the banker for the particular deal net~ork.~

clusion did not alter the strength of the other co- efficients. This variable was also not significant in any of the equations involving outcome, and its presence either had no impact on the effects of our hypothesized predictors or (in three cases) increased them.

It is possible that bankers may have entered into a deal with the idea that the deal was either particularly "strong" or "weak," and this notion may have affected the character of the networks the bankers constructed. If this were the case,

Tables la and Ib present means, standard deviations, and correlations among the vari- ables in our analysis. We present these in two separate tables because only a subset of our cases include approval networks. Among the 194 deals for which we have outcome in- formation, only 151 (77.8 percent) pro- ceeded far enough to involve an approval network. The remaining 43 deals failed be- fore reaching the approval process. Of the 151 deals that reached the approval process, 96 (63.6 percent) were successfully closed. Table la includes descriptive statistics for and correlations among our variables for the 173 deals for which we have complete data, regardless of whether the deals reached the approval stage. Table lb includes the same

then an observed association between network structure and deal closure might be a spurious consequence of the banker's initial view of the deal. There are two reasons that this possibility is unlikely to affect our findings. First, based on our interviews, this process appears to be rela- tively rare. We observed three deals (among more than 200) that were, from the start, viewed as "blockbusters," in which a large number of bankers sought to give their support. Except for these three cases, bankers did not speak of deals in terms of their strength, but rather in terms of their degree of complexity or risk. Second, in each of the three "blockbuster" cases, the most prominent outcome was the presence of a large approval network. By controlling for network size, we increase the likelihood that this process did not distort the effect of network density. It is also possible that the character of competing deals from other banks may have affected whether UniBank successfully closed a deal. Our interviews suggest that the bank faced competi- tion from other banks in most of its deals, includ- ing those involving even long-term customers. We saw little evidence from our interviews to in- dicate that competition was systematically asso- ciated with any of our predictors, however. We also asked bankers, in response to failed deals, an open-ended question regarding why they thought the deal failed. Losing out to a competi- tor was mentioned in about 20 percent of the cases-the most common responses were "cus- tomer's needs changed" and "customer's situation changed." There was no discernable pattern between reason for failure and any of our predic-

tors of closure.

Variable     Mean     S.D.     (1)     (2)     (3)     (4)     (5)     (6)     (7)     (8)     (9)     (10)     (11)     (12)     (13)     (14)     (15)
(I) Outcome of the deal (1 =closed)     .636     .483     1.00     -.I2     .09     -.I3     .I2     .03     .03     .02     -.I4     -.I5     .08     ,413     -.32     -.39     -.20
(2) Economic uncertainty (In)     11.549     3.161     -    1.00     -.06     -.06     -.02     -.07     .04     -.03     .24     .04     -.08     -.07     .19     .I7     -.08
(3) Customer trust (4) Complexity of deal (5) Consensus (6) Banker rank (7) Lead bank (8) Confidence in management     .510 3.363 3.885 2.338 .987 3.570     ,502 1.107 1.01 1 1.006 ,792 ,579     --1.00 .13 .24 .18 .I2 ---1.00 -.I3 .27 .12 1.00 .O1 .06 1.00 .O1 1.00 ----------------------    .34 .12 .08 .I2 .05 1.00     .O1 -.I0 .I8 .24 .02 .O1     -.05 .01 .02 .I0 .OO -.00     .O1 .05 -.05 -.I2 .06 .04     .08 .17 .12 .13 .O1 -.09     .OO -.09 .01 .19 .06 .05     -.06 -.09 -.I9 .O1 .07 .O1     -.I 1 -.01 -.03 -.25 -.02 -.03     0 5 r Z m 20 B > 2 u
(9) Information network strength (10) Information network density (1 1) Information network hierarchy (12) Approval network size (1 3) Approval network strength (14) Approval network density     ,677 ,577 ,324 3.662 ,702 ,656     ,161 .207 ,253 2.094 ,171 ,209     --------1.00 .58 -.23 -.I2 .67 -1.00 -.I3 -.I5 .40 1.00 .04 -.10 1.00 -.30 1.00 ------------------------------------------------------    .41 .51 -.05 -.51 .70 1 .OO     -.28 -.32 .29 -.17 -.30 -.20     Z x u m2 !!! 0 2 i:> E 2G1

--------------1 .OO

(15) Approval network hierarchy .375 ,314



Note: Number of deals =137. (D

information for the 137 deals, for which we have complete data, that reached the ap- proval stage.7

Except for the correlations between alter network density and the controls for net- work size and ego network strength, none of the correlations among the predictors is large enough to suggest concern about multicollinearity. Note that the level of eco- nomic uncertainty and the complexity of the deal are not correlated, both among all deals and among deals that reached the ap- proval stage. This suggests that these two variables are measuring distinctly different dimensions of uncertainty. The positive (al- beit small) association between complexity and customer trust suggests, not surpris- ingly, that bankers were more willing to pursue a complex deal when they had had positive prior relations with the customer.

Because we examine more than one deal per banker, our observations are not statisti- cally independent. In the analyses that fol- low, we use a technique to compute robust variance estimates for clustered observations in which we transform the variance-covari- ance matrix of the regression coefficients to take into account the nonindependence of deals within individual bankers. The stan- dard errors reported in Tables 2 and 3 are based on these estimates. A description of the model is presented in Appendix A.

Table 2 presents the tests of Hypotheses 1 and 2. In Model 1, we examine the effects of our two hypothesized indicators of uncer- tainty-the natural logarithm of the product of exposure and company credit rating (eco- nomic uncertainty) and degree of trust of the customer firm-on the strength of the bank- ers' relations with members of their informa- tion network.

Model 1 reveals that the level of economic uncertainty has a statistically significant positive association with the strength of the banker's ties with those whom he or she con-

'The effect of size of the information network on closure was virtually zero, and it had no ef- fect on the strength of the other coefficients. Be- cause of the null result and the fact that we did not develop an explicit hypothesis for the effect of information network structure on closure, we omit this variable from the correlation matrices and the equations we present below.

sults for information related to a deal. This finding is consistent with Hypothesis 1: When uncertainty is high, bankers tend to turn to those with whom they are close. The level of trust of the customer firm was not associated with the strength of a banker's in- formation network ties, a finding inconsis- tent with Hypothesis 1. And, although we did not hypothesize a specific effect of com- plexity in Hypothesis 1, this variable also was not associated with strength of ties. One possible reason for this is that in some com- plicated deals, bankers may be forced by the nature of the deal to interact with a wide range of others. Support for this notion is found in a separate analysis (available on re- quest) in which we used size of the informa- tion network rather than tie strength as the dependent variable. Complexity of the deal was the only significant predictor in the equation, suggesting that in complex deals bankers will consult with a relatively large number of colleagues. Economic uncertainty had no effect on the size of the information network. This result suggests, consistent with our argument, that bankers turn to their strong ties, rather than to a large number of colleagues, for information in high-risk deals.8 The fact that the economic uncer- tainty effect on tie strength holds even when we control for the complexity of the deal supports our contention that under condi- tions of high risk, bankers will turn for ad- vice to those they trust.

The equation predicting the strength of ties with members of a banker's approval network (Model 3 of Table 2) requires some discussion. The fact that approximately 22 percent of our deals failed before reaching the approval process means that the deals for which we have data on approval networks may constitute a biased sample of the total number of transactions. This phenomenon, known as sample selection bias, is a com- mon problem in the analysis of social sci- ence data (Berk 1983; Winship and Mare 1992). The most widely used approach to

To further examine this view, we inserted size of the information network as a control into the equation predicting tie strength. This variable was not significantly associated with tie strength, nor did its presence have any effect on the strength of the economic uncertainty coefficient.

Table 2. Unstandardized Coefficients Predicting the Effects of Uncertainty and Other Independent Variables on Strength of Banker's Ego Networks

Independent Variables Constant

Economic uncertainty (In) Customer trust Complexity of deal Consensus Banker rank Lead bank Confidence in management

Number of deals

Information Selection Approval Network Equation Network Model la Model 2b Model 3a

.007+ (.004)


173 173 137

Note: Numbers in parentheses are standard errors based on robust variance estimates with clustering. The dependent variable for Model 1 is strength of information network ties, regardless of whether the deal reached the approval stage. The dependent variable for Model 3 is strength of approval network ties. The dependent variable for Model 2 predicts whether the deal reaches the approval process.

a OLS models.

Probit model.

*p< .05 **p< .01 ***p< .001 (two-tailed tests)

+p < .05 ++p< .O1 +++p< ,001 (one-tailed tests)

handling problems of sample selection bias is a two-stage model developed by Heckman (1979). In the standard Heckman model, the investigator first computes a probit model predicting factors that affect the probability of being selected into the outcome condition (in our case, the probability of having an ap- proval network). From the selection equa- tion, the researcher then uses the probit co- efficients to estimate, for each observation, a value (A)that represents a hazard rate (the probability that a from the sample, that is, fai1 to reach the approval process, conditional on being at risk of dis- appearing) (Berk 1983:390-9 1 ; Greene

1995:638-40). The A is then inserted as a variable into the substantive equation.

To compute this model, we must identify a set of predictors for the selection variable (existence of an approval network). Al- though it is not absolutely necessary, esti- mation is facilitated if we ensure that there is at least one variable in the selection equation that does not appear in the sub- stantive eq~ation.~

For our predictors of a

9 Identical selection and substantive equations increase the likelihood of multicollinearity be- tween A and the predictors in the substantive equation (Berk 1983:396-97).

deal reaching the approval process, we se- lected seven variables: the six that served as predictors of strength of ties in the informa- tion network, and the banker's reported confidence in the customer firm manage- ment's ability to keep the firm on a strong financial footing (on a 1 to 4 scale). The six variables from Model 1 of Table 2 were chosen to render the analysis of the strength of approval network ties as close as pos- sible to the analysis of the strength of infor- mation network ties.

Model 2 of Table 2 presents the selection equation predicting the existence of an ap- proval network. Interestingly, only one of the seven predictors in the selection equation is significant using a two-tailed test, al- though a second variable approaches signifi- cance. This latter variable, trust of the cus- tomer, is (as we would expect) positively as- sociated with the presence of an approval network, although its coefficient falls slightly short of statistical significance (p = .06). Contrary to what we might expect, however, high levels of economic uncertainty make it more likely that a deal will reach the approval process. Although this finding appears counterintuitive, our inter- views suggest a straightforward explanation. Several bankers told us that for high-risk deals it is important to begin securing ap- proval as early as possible. For low-risk deals, in which the securing of approval is not problematic, bankers will often refrain from seeking approval until they believe the deal has a good chance of closing. This means that a significant number of low-risk deals fall apart before the banker has estab- lished an approval network. As a conse- quence, high-risk deals have a higher prob- ability of reaching the approval process.

Moving to the substantive analysis, be- cause the Heckman model produces ineffi- cient estimates, some authors (Berk and Ray 1982:382; Breen 1996:40) recommend the use of an alternative, maximum likelihood, estimator. Breen (1996:70-71) describes a test for whether the maximum likelihood es- timator, as compared with OLS, should be used. This involves regressing the hazard rate (A)on the variables in the substantive equation. If the coefficient of determination from this equation is close to zero, Breen recommends the use of OLS. We conducted this test and found that the estimated R2 was zero, rounded to six places. All seven regres- sion coefficients, plus the constant, had t-statistics of virtually zero. Consistent with Breen's suggestion, the results of the equa- tions using OLS and maximum likelihood estimators were virtually identical. This finding suggests that OLS estimates are ap- propriate.

Model 3 presents the OLS equation pre- dicting the strength of the bankers' relations with members of their approval networks. The findings in this model are consistent with the first equation's finding on informa- tion networks: In deals with high levels of economic uncertainty, bankers are more likely to turn to colleagues with whom they are close. As in the information network model, the effect of trust of the customer is not significant. Although we did not issue an explicit hypothesis for the effect of com- plexity, we note that this effect approaches statistical significance in a two-tailed test (p = .08). This finding raises the possibility that bankers are using larger networks for complex deals. A separate analysis (not shown) provides support for this notion. Complexity has a significant positive effect on the size of the banker's approval net- work. As with the information network, the use of larger approval networks occurs in complex deals but not in risky ones. Eco- nomic uncertainty does not increase the size of bankers' approval networks; in fact, the effect of uncertainty on network size is negative. The findings in Table 2 thus indi- cate that bankers turn to their strong-tie col- leagues for support in high-risk situations, but the level of economic uncertainty has no effect on the size of the approval net- work.1° In contrast, complexity leads to the creation of larger networks but does not af- fect the strength of the bankers' ties. The banker's trust of the customer firm does not predict the nature of the ties the banker con- sults, but the findings on the role of eco-

lo As with our analysis of strength of the bank- ers' information network ties, we inserted size of the approval network into the equation for Model

3. In this case, the effect was significantly nega- tive: Large networks were associated with weaker ties. The presence of this variable had no impact on the size or strength of the economic uncertainty coefficient, however.

nomic uncertainty are consistent with Hy- potheses 1 and 2."


Like our analysis of the determinants of ap- proval network tie strength, our analysis of deal closure requires the consideration of a sample selection model. An approach similar to the Heckman model is available for situa- tions in which both the selection and substan- tive equations contain binary dependent vari- ables. In this case, the substantive equation is a second probit model, with closure as the dependent variable. An alternative approach also exists under these conditions, however. If both dependent variables are dichotomous, it is possible to compute the sample selection model with a bivariate probit design (Greene 1995:465). In this approach, the two probit

l1 As noted above, our two-stage model treats ego network strength as endogenous in the first stage and alter network density as exogenous in the second stage. To what extent does uncertainty predict the structure of bankers' alter networks? To examine this, we substituted alter network density for ego network strength as our depen- dent variables and recomputed the equations in Table 2. None of our predictors was significantly associated with information network density. This means that although bankers turned to their strong direct ties for information in high-risk deals, they did not appear to strategically con- struct either a dense or a sparse alter network. For approval network density, on the other hand, the results were almost identical to those predicting approval ego network strength. High levels of economic uncertainty were associated with high levels of alter network density. This is partly an artifact of the high correlation between tie strength and density; when we control for tie strength, the economic uncertainty effect disap- pears. Another possible explanation is that bank- ers sometimes may be more successful in gain- ing support from a superior if they have already enlisted the support of another colleague whom the superior trusts. This is consistent with the ex- ample of the approval network from our inter- views, described above. The fact that the finding disappears once we control for tie strength sug- gests caution with this interpretation, however. These findings, along with data from our inter- views, provide little systematic evidence that bankers were strategically tailoring their net- works to take into account relations among their alters.

models are computed simultaneously through an iterative procedure. The error terms for the two equations are correlated, creating an autocorrelation estimate, rho (p), which is then included in the model.

Although Greene (1995:646) recommends the bivariate probit model, when we applied this model to our data, our log-likelihood function did not converge to a clear solution. We therefore decided to examine Hypoth- eses 3 through 5 using both the two-stage and bivariate probit approaches. With a few exceptions, the two models yielded virtually identical results. Because there were differ- ences, and because we lack a clear basis on which model to choose, we report both sets of results in the analyses that follow.12

Because we are using the selection equa- tion for the existence of the approval net- work (Model 2 of Table 2), we report only the substantive equations for deal closure in Table 3. Selection equations in the bivariate probit analysis vary across models, although they tend to be similar. We have included the associated bivariate probit selection equa- tions in Appendix B. Models la and lb in Table 3 present our basic model, which tests Hypotheses 3 through 5 as well as the effects of the density and hierarchy of the bankers' information networks. Model la provides the two-step estimates, while Model lb pro- vides the bivariate probit estimates.

Hypotheses 3 through 5 predict the effect of three variables on the likelihood of a deal successfully closing: uncertainty, density of the approval network, and hierarchy of the approval network. All three variables are predicted to be negatively associated with closure. The first uncertainty variable-eco- nomic uncertainty (the log of the quantity exposure times the customer credit rating)- is not significantly associated with the prob- ability of closure in Models la or lb. This variable was significantly negatively associ- ated with closure in the two-stage model when complexity of the deal was removed from the equation (not shown). The second uncertainty variable-the banker's trust of

l2 Unlike the analysis of approval network tie strength in Table 2, a regression of 1on the pre- dictors in the outcome equation indicated the need to include 1in all of the two-stage equa- tions in Table 3.

Table 3. Unstandardized Coefficients Showing the Effects of Uncertainty and Network Characteristics on the Likelihood of Deal Closure

Two-Stage Sample Selection Models and Bivariate Probit Modelsa
Independent     Model     Model     Model     Model     Model     Model     Model     Model
Variables     la     lb     2a     2b     3 a     3b     4a     4b
Constant     2.063     2.353*     -.310     1.813     2.989     3.102*     ,626     2.281
    (2.824)    (1.072)     (2.931)     (1.033)     (3.006)     (1.243)     (3.088)     (1.176)
Economic     ,050                             
uncertainty (In)     (.105)                             
Customer trust     .597                             
    (.5 16)
Complexity of deal     -.462+
Consensus     ,052
Banker rank     ,101
Lead bank     .3 1 1
Information Network     
Density     -.329     -.274     -.453     -.472     -.322     -.311     -.468     -.469
    (.656)    (.636)     (.683)     (.665)     (.678)     (.657)     (.734)     (.692)
Hierarchy     1.225*     ,938     ,973     .882     1.403*     ,964     1.134     .826
    (.600)    (.636)     (.621)     (.603)     (.641)     (.562)     (.664)     (.544)
Approval Network     
Density     -3.926+++ -2.728+++ -3.009+++ -2.358+++ -2.306++ -1.616+     -1.369     -1.174
    (.789)    (.689)     (.925)     (.762)     (.917)     (.926)     (1.058)     (1.096)
Hierarchy     -2.070+++ -1.544++ -1.675+++ -1.412++ -2.616+++ -1.91 1++ -2.227+++ -1.758++
    (.484)    (.582)     (.522)     (.577)     (.548)     (.675)     (.569)     (.621)
Size     -    -    .228++     .148+     -    -    .222++     .151+
            (.084)    (.073)             (.078)     (.079)
Strength     -    -    -    -    -3.194++ -2.150+     -3.184++ -2.052
                    (1.166)    (1.246)     (1.172)     (1.391)
Lambda (a)     2.619     -    3.749     -    2.763     -    3.860     -
    (2.859)        (2.828)         (2.957)         (2.962)     
Rho (PI     -    .999*** (.158)     -    ,999"' (.091)     -    .998*** (.183)     -    .999*** (.136)
x2     47.849***     -    55.402***     -    53.774***     -    61.006***     -
d.f.     11     -    12     -    12     -    13     -
Log likelihood     --146.184     --142.243     --143.525     --140.193

Note: Numbers in parentheses are standard errors based on robust variance estimates with clustering. a Models la, 2a, 3a, and 4a are two-stage models. Models lb, 2b, 3b, and 4b are bivariate probit models.

Number of deals for all models = 136. *p< .05 **p< .O1 ***p< ,001 (two-tailed tests) +p< .05 ++p< .O1 +++p< ,001 (one-tailed tests)

the customer firm-is not statistically sig- nificant in the two-stage model in Model la, but it does reach statistical significance in the bivariate probit model (Model lb). The effect of deal complexity is strongly signifi- cantly negative in both equations: The more complex the deal, the less likely it is to close. These findings provide mixed support for Hypothesis 3.

The fact that economic uncertainty fails to predict closure seems counterintuitive: We initially expected risky deals to be more dif- ficult to accomplish. One possible explana- tion for this finding may be the restructuring of deals that occurs during the approval pro- cess. We learned during our interviews that it is the bank's policy not to turn deals away, but to restructure them until the bank be- lieves it will secure an acceptable rate of re- turn. Such a reconfiguration might involve securing an initial fee, charging a higher in- terest rate, or selling off parts of a loan. In such situations, the banker benefits from the networks that produce creatively restruc- tured deals, ones that can satisfy both the customer and the bank.

Among the remaining variables, the banker's report of the degree of consensus within the bank around the deal was not as- sociated with success of the deal in either of the models." Although the coefficients for banker rank and having UniBank as a lead bank are positive, neither effect reaches sta- tistical significance. The fact that rank is not significantly associated with closure may be in part because highly ranked bankers tend to work on more complex deals.

Consistent with our earlier discussion, the density of the bankers' information networks was not significantly associated with the likelihood of closure. The information that bankers receive from colleagues may sug- gest that the deal is not worth pursuing, or that it should be handled differently (which may lead to the banker abandoning the deal and starting anew). The hierarchy of the bankers' information networks was posi-

l3 One possible reason for this result may be a lack of validity of our indicator. Some bankers assumed, for example, that once a deal was agreed upon by all parties in the bank, it had a high degree of consensus, even if there had been considerable controversy at earlier stages.

tively associated with the likelihood of clo- sure in the two-step model but was not sig- nificant (in a two-tailed test) in the bivariate probit equation. To the extent that this effect is positive, one possible explanation, sup- ported by our interviews, is that bankers of- ten reported that it was important to have good relations with product specialists when putting together a deal. Several bankers sug- gested that a strong relationship with a single product specialist may be sufficient to produce a successful deal.

The effects of approval network density and hierarchy are strongly negative in both versions of Model 1, providing support for Hypotheses 4 and 5. The broader the range of colleagues whose support is enlisted, the more likely a deal is to close.14 Even when we control for the density of the approval network, the level of hierarchy is still strongly associated with closure. Bankers whose approval networks are dominated by one or a small number of alters are less likely to successfully close their deals.

The findings in Models 1 a and 1 b of Table 3 thus suggest strong support for Hypotheses 4 and 5 and partial support for Hypothesis 3. Before we assume unambiguous support for Hypotheses 4 and 5, however, we must ad- dress two remaining issues: the size of the approval network and the strength of the banker's direct relations. Both of these vari- ables, especially ego network strength, are highly correlated with approval network den- sity, and the presence of all three variables in the same equation raises concerns about multicollinearity (the coefficient of determi- nation of density regressed on the remaining independent variables is .70).Our solution is to enter each of the two controls separately

l4 One possible implication of our argument is that sparse networks might be especially valuable in high-uncertaintv deals. If so. this would sup- gest the presence of interaction effects between density and uncertainty. To examine this, we cre- ated interactions combining approval network density with economic uncertainty and complex- ity, bdth of which we would expect to be nega- tive. We did observe negative coefficients for these interactions, but none was statistically sig- nificant. We also examined interactions involv- ing density and trust, all of which were also non- significant. The effect of density appears to hold regardless of the level of uncertainty.

and examine whether their presence affects the strength of the density effect. Models 2a and 2b of Table 3 include network size. In Models 3a and 3b we remove network size and insert ego network strength. The inser- tion of approval network size has several in- teresting consequences. The effect of cus- tomer trust becomes more strongly signifi- cant in the bivariate probit model (Model 2b) and approaches significance (p = .06) in the two-stage model (Model 2a). The effect of information network hierarchy drops below significance in the two-stage model, and the effect of complexity increases slightly in both models. If our earlier argument about the value of a broad range of support is accu- rate, then we would expect a large approval network to be consistent with success of the deal. We find a positive effect of approval network size in both models. Even when we take this variable into account, however, the effect of network density remains strongly negative, as does the effect of hierarchy. Clearly the effect of density is not an artifact of network size.

A similar finding arises when we control for ego network strength (Models 3a and 3b). Although the presence of this variable does reduce the effect of density, especially in the bivariate probit model, both effects remain significant in the predicted negative direction, despite the possible presence of multicollinearity (the squared multiple cor- relation of alter density on the remaining re- gressors, including ego network strength but excluding network size, is .62). The strong negative effect of approval network hierar- chy holds in both equations, as does the negative effect of complexity. The positive effect of customer trust drops just barely be- low statistical significance (p = .0506) in the bivariate probit model and remains nonsig- nificant in the two-stage model.

The important findings in these two sets of equations, however, are that the effects of approval network density and hierarchy re- main significantly negative in all four equa- tions, even when we control for the effects of network size and ego network strength. These results provide strong support for Hy- potheses 4 and 5.

Despite the likely presence of multicol- linearity, we also examined models with ap- proval network size, strength, and density included simultaneously. These are pre- sented in Models 4a and 4b of Table 3. The simultaneous inclusion of network size and strength depresses the density effect below the boundary of statistical significance. The coefficients for strength and size remain sig- nificant in the expected direction, as does the effect of network hierarchy, although the ef- fect of strength drops just below significance (p = .07) in the bivariate probit model (the simultaneous insertion of network density and tie strength into Model 4b reduces the log likelihood function by 9.98, which is equivalent to an improvement x2 of 19.96 with 2 d.f., p < .001).

We believe that this finding on approval network density does not damage the sup- port for Hypothesis 4, for two reasons. First, although multicollinearity among approval network size, strength, and density may lead to inflated standard errors for all three coef- ficients, this is especially likely to be the case for the density coefficient because the correlation between size and strength is only -.30, while that between size and density is -.51 and that between strength and density is .70. Second, even without a significant density effect, the combined findings for the size and strength variables are consistent with our argument. Network theory suggests that a large number of weak ties will create a diverse network, which is precisely the ba- sis for our hypothesized effect of low den- sity. It is true that strong ties do not neces- sarily form a dense alter network, nor do weak ties necessarily form a sparse one. Ex- isting theory and our own findings indicate that this is likely to be the case, however. Given the imperfect nature of our measures, the combination of large size and weak ties in a banker's network would seem to be a reasonable alternative indicator of the con- cept-diversity of perspectives-that our network density measure is designed to tap.15 The findings in Models 4a and 4b, along with those from the other equations in Table 3, are thus consistent with our argu- ment in Hypothesis 4.

l5 This holds, we believe, even if we make al- lowances for the possibility that some of the size effect could be capturing the degree of risk in the deal. Note, for example, that neither economic un- certainty nor trust is associated with network size.

The results provide considerable support for four of our five hypotheses, and mixed sup- port for the fifth. Under conditions of high economic uncertainty, bankers are more likely to consult with strong-tie associates for both information and support, consistent with an important component of Hypotheses 1 and 2. Economic uncertainty does not have a significant partial effect on the closure of a deal, as we suggested in Hypothesis 3, but our finding that low levels of complexity are associated with closure is consistent with the hypothesis, and our findings on customer trust indicate that in at least some of our models, trust is associated with closure. Most significant, we find strong support for Hypotheses 4 and 5: Relatively sparse and nonhierarchical approval networks are con- ducive to successful closure of a deal.

These findings suggest a paradox, how- ever. On one hand, when uncertainty is high, bankers cling to those they trust, with whom they are closely tied. On the other hand, embeddedness in strongly connected net- works is precisely the condition that makes it more difficult to close deals. Uncertainty cre- ates conditions that trigger a desire for the familiar, and bankers respond to this by turn- ing to those with whom they are close. Yet it is this very action that makes it more diffi- cult for the banker to be successful. Not only does this illustrate of the simultaneous weak- ness of strong ties and the strength of weak ties, but it also shows how our social instincts can run counter to our best interests.

Our findings touch on a number of issues. They support the contention from much or- ganizational analysis that actors will use trusted individuals or symbols to deal with uncertainty, a finding consistent with a range of classic (Granovetter 1974; Kanter 1977; Spence 1974) and more recent studies (Galaskiewicz, Dowell, and Bielefeld 2001; Podolny 1994; Stuart, Hoang, and Hybels 1999). Social networks have been found to be important means of managing uncertainty in other organizational contexts, and they are significant in the banking world as well.

Yet it is not only the presence or use of networks but also their specific character that is important. Granovetter's (1973) strength-of-weak-ties hypothesis and Burt's (1992) concept of structural holes are both relevant to our study, and both views empha- size the importance of distinguishing among types of ties and types of networks. In Granovetter's model, an actor is likely to re- ceive a greater volume of information from weak ties. In Burt's model, an occupant of a structural hole is able to control the flow of information between different groups. The mechanism created by sparse approval net- works among our bankers differs slightly from both of these formulations. We suggest that what a sparse approval network creates is a diversity of views and potential criti- cisms that compel the banker to create a higher quality product.

Consider as an analogy a scholar present- ing an argument to a group of like-minded peers. Although the peers may be knowledge- able about the topic and therefore able to pro- vide many helpful criticisms, they are also likely to share many broad assumptions with the presenter and with one another. If the scholar then presents his or her material to a different audience, to whose criticisms the scholar has not been previously exposed, he or she may have a more difficult time con- vincing his or her listeners. On the other hand, had the scholar originally presented his or her work to an audience of unconnected alters (a group that is likely to have a more diverse set of views), he or she may be better prepared to anticipate the criticisms of a wider range of audiences in the future. We argue that a similar process occurs among our bankers. A deal that receives support from a tightly connected group of alters may receive less probing criticism (or at least a less broad set of criticisms) from the banker's colleagues. We believe that, ceteris paribus, these deals will be less attractive to the cus- tomer than will a deal that has been subjected to feedback from a more diverse group of colleagues. Another example of this phenom- enon is illustrated in David Halberstam's The Best and the Brightest ([I9721 1992), in which a group of brilliant, but like-minded,

I presidential advisors uncritically ushered the

I United States into full-scale military involve- ment in Vietnam. In addition to the acquisi-

I tion of information (the weak-tie hypothesis) and the ability to manage the flow of infor- mation (the structural-hole hypothesis), then, our analysis suggests a third potential ben-

efit of sparse social networks: criticisms that allow an actor to anticipate a variety of con- tingencies, or what we might call the multiple-lens hypothesis.

Our finding also provides an example of one of the most foundational concepts in the social sciences: the unanticipated consequences of purposive social action (Merton 1936).Sociologists since Durkheim have fo- cused on the nonrational elements of eco- nomic behavior, emphasizing the role of symbols, rituals, and trust. Our focus has been on bankers operating in a high-stakes arena in which there are strong incentives and measurable levels of success, exactly the conditions that even sociologists would con- cede lend themselves to rational behavior. We have strong reason to believe that our bankers are intendedly rational. They cer- tainly have a clear set of goals and a set of strategies to accomplish them. The bankers are constrained by bounded rationality, how- ever, in that they operate with a lack of clear information. Thus, even when they engage in purposive, intendedly rational behavior, their decisions may have consequences the opposite of what they expect. We have un- covered one such strategic paradox: The use of strong ties seems to be a rational strategy for managing uncertainty, yet if bankers' ul- timate goal is to close deals, then the use of strong ties appears counterproductive. It is the sparsely connected, diverse approval net- works that are most closely associated with the successful closure of deals.

This finding raises a series of questions that will require further work to address: How aware are these bankers of the rationale behind their use of approval networks and the consequences of their choices? Several bank- ers with whom we spoke after the comple- tion of our formal interviews noted that our argument about the value of sparse approval

The Analytic Model

Because we examined more than one transaction per banker, our 194 deals are potentially nonindepen- dent. There are several ways to handle this problem. One is to reduce the degrees of freedom in our sta- tistical tests to reflect the number of independent observations. Another approach is to scale values on transactions to the person-means, a variant of what

networks "rang true." They did not appear to be aware of the negative association between the strength of their individual ties and the sparseness of their network, however. We also do not know the extent to which the use of strong ties actually reduced bankers' uncertainty. Nor do we know the extent to which bankers are aware of the potentially contradictory consequences of their deci- sions. And although we know that bankers have discretion in their use of approval net- works, we lack information on the variation in the amount of choice that bankers had across deals. All of these issues could be ad- dressed in follow-up interviews. They would provide a unique opportunity to examine the creation, and strategic use, of social net- works.

Certainly we view our results as prelimi- nary, and there are unique characteristics of deals that cannot be fully understood in the aggregate form presented here. We believe, however, that we have uncovered a genuine strategic paradox, one that could have real consequences for the success of organiza- tions and the individuals within them.

Mark S. Mizruchi is Professor of Sociology and Business Administration at the University of Michigan. His work focuses on the areas of eco- nomic, organizational, and political sociology. His current projects include a study of the glo- balization of American banking and an examina- tion of the diffusion and uses of knowledge in the sociology of organizations.

Linda Brewster Stearns is Professor of Sociol- ogy at the University of California, Riverside. Her research interests focus on economic and politi- cal sociology, with a current concentration on networks andfinancial decision-making, environ- mental policymaking, and merger waves. She is currently writing a book that examines the deter- minants and consequences of merger wave activ- ity in the United States from 1890 to the present.

is called least squares with dummy variables (LSDV). Hannan and Young (1977) have shown that LSDV estimates are superior to OLS estimates in terms of both consistency and efficiency. Be- cause none of our bankers had more than three deals, and for many we have information on only one or two, the LSDV approach is not feasible. An alternative is to compute robust estimates of the variance-covariance matrix, adjusted for clustered observations. This approach is well-suited to situa- tions such as ours in which the number of "groups" relative to observations within the groups is large (Rogers 1993). The use of robust standard errors was developed by Huber (1967) and White (1980). The adjustment for clustering was developed by Rogers (1993). The computer program Stata has a moduleto compute robust standard errors with clus- tering. Because we were conducting our data analy- sis with LIMDEP, Professor William Greene, the author of the program, generously wrote an algo- rithm for us to perform the computation within LIMDEP. This algorithm is based on the same prin- ciples as those described in the Stata manual (Greene, private communication; Stata 1999:256- 60). The cluster estimate of the asymptotic vari- ance-covariance matrix can be written as

where V is the standard variance-covariance matrix of the regression coefficients [O (XIX)-I], and G is an n x k matrix of sums of the individual scores (the first derivatives of the log likelihood) for the observations in each cluster, where n equals the number of clusters (in our case, the number of indi- vidual bankers), and k is the number of exogenous

variables plus one (the constant). For OLS regres- sion, the elements in G are

where e, is the OLS residual of the observation t of group i, and xi,is the value of the independent vari- able for the particular observation. For the probit model, the matrix equation is the same, but the ele- ments of G are

gi =C(aitxit),

where ,Ii,is the inverse Mills ratio from the probit model for the particular observation (Greene, pri- vate communication; Greene 1995:640).

Greene's LIMDEP algorithm operates by first computing a standard OLS or probit model, assum- ing independence, and then storing the residuals or inverse Mills ratios as input for the computation. The regression coefficients of the standard and clus- tering models are identical; only the standard errors are different. The standard errors from our robust cluster analyses were in every case similar to the conventional standard errors, and the substantive conclusions derived from them were identical. Re- sults of the analyses with the conventional standard errors are available on request.

Probit Coefficients from Selection Equations Predicting Whether a Deal Will Reach the Approval Stage: Bivariate Probit Models in Table 3

Independent Variables Model Ib Model 2b Model 3b Model 4b Constant

Economic uncertainty (In) Customer trust Complexity of deal Consensus Banker rank Lead bank Confidence in management

Note: Each equation contains probit coefficients (with standard errors, based on robust variance estimates with clustering, in parentheses) corresponding to the bivariate probit model of the corresponding model number pre- sented in Table 3. Number of deals = 173 for all models.

*p < .05 **p< .O1 ***p< ,001 (two-tailed tests)

Allison, Paul D. 1978. "Measures of Inequality." American Sociological Review 43:865-80.

Baker, Wayne E. and Robert R. Faulkner. 1993. "The Social Organization of Conspiracy: Ille- gal Networks in the Heavy Electrical Equip- ment Industry." American Sociological Review 58:837-60.

Berk, Richard A. 1983. "An Introduction to Sample Selection Bias in Sociological Data." American Sociological Review 48:386-98.

Berk, Richard A. and Subhash C. Ray. 1982. "Se- lection Biases in Sociological Data." Social Science Research 11 :352-98.

Breen, Richard. 1996. Regression Models: Cen- sored, Sample Selected, or Truncated Data.

Thousand Oaks, CA: Sage. Burt, Ronald S. 1992. Structural Holes. Cambridge, MA: Harvard University Press.

. 2000. "The Network Structure of Social Capital." Research in Organizational Behavior 22:345-423.

Davis, Gerald F. 1991. "Agents without Prin- ciples? The Spread of the Poison Pill through the Intercorporate Network." Administrative Science Quarterly 36:583-6 13.

DiMaggio, Paul J. and Walter W. Powell. 1983. "The Iron Cage Revisited: Institutional Iso- morphism and Collective Rationality in Orga- nizational Fields." American Sociological Re- view 48: 147-60.

Galaskiewicz, Joseph. 1985. Social Organization of an Urban Grants Economy. Orlando, FL: Academic Press. Galaskiewicz, Joseph, Myron Dowell, and Wolfgang Bielefeld. 2001. "Testing Ecologi- cal and Social Capital Theories of Organiza- tional Growth." Paper presented at the Stanford Strategic Management Conference, March 3, Stanford University, Stanford, CA.

Gargiulo, Martin and Mario Benassi. 2000. "Trapped in Your Own Net? Network Cohe- sion, Structural Holes, and the Adaptation of Social Capital." Organization Science 11: 183-96.

Granovetter, Mark. 1973. "The Strength of Weak Ties." American Journal of Sociology 78: 1360-80. . 1974. Getting A Job. Cambridge, MA: Harvard University Press. . 1985. "Economic Action and Social Structure: The Problem of Embeddedness." American Journal of Sociology 9 1 :481-5 10.

Greene, William H. 1995. LIMDEP. Version 7.0. Bellport, NY: Econometric Software.

Gulati, Ranjay. 1998. "Alliances and Networks." Strategic Management Journal 19:293-3 17. Halberstam, David. [I9721 1992. The Best and the Brightest. New York: Random House.

Hannan, Michael T. and Alice A. Young. 1977.

"Estimation in Panel Models: Results on Pool- ing Cross-Sections and Time-Series." Sociological Methodology 752-83.

Heckman, James J. 1979. "Sample Selection Bias as a Specification Error." Econometrica 45: 153-61.

Huber, Peter J. 1967. "The Behavior of Maxi- mum Likelihood Estimates under Non-stan- dard Conditions." Proceedings of the Fifth Berkeley Symposium on Mathematical Statis- tics and Probability 1 :22 1-33.

Kanter, Rosabeth Moss. 1977. Men and Women of the Corporation. New York: Basic.

Labianca, Giuseppe L., Daniel J. Brass, and Bar- bara Gray. 1998. "Social Networks and Per- ceptions of Intergroup Conflict: The Role of Negative Relationships and Third Parties." Academy of Management Journal 4155-67.

Lin, Nan. 2000. "Inequality in Social Capital." Contemporary Sociology 29:785-95. March, James G. 1994. A Primer on Decision Making. New York: Free Press.

Merton, Robert K. 1936. "The Unanticipated Consequences of Purposive Social Action." American Sociological Review 1 :894-904.

Meyer, John W. and Brian Rowan. 1977. "Insti- tutionalized Organizations: Formal Structure as Myth and Ceremony." American Journal of Sociology 83:440-63.

Mizruchi, Mark S. 1992. The Structure of Cor- porate Political Action. Cambridge, MA: Harvard University Press.

Pfeffer, Jeffrey and Gerald R. Salancik. 1978. The External Control of Organizations. New York: Harper and Row.

Podolny, Joel M. 1994. "Market Uncertainty and the Social Character of Economic Exchange." Administrative Science Quarterly 39:458-83.

Powell, Walter W., Kenneth W. Koput, and Lau- rel Smith-Doerr. 1996. "Interorganizational Collaboration and the Locus of Innovation: Networks of Learning in Biotechnology." Administrative Science Quarterly 41:1 16-45.

Rogers, William H. 1993. "Regression Standard Errors in Clustered Samples." Stata Technical Bulletin 13: 19-23.

Simon, Herbert A. 1947. Administrative Behav- ior. New York: Macmillan.

Spence, A. Michael. 1974. Market Signaling.
Cambridge, MA: Harvard University Press.
Stata. 1999. Stata User's Guide. Release 6. Col-

lege Station, TX: Stata Press.

Stearns, Linda Brewster and Kenneth D. Allan. 1996. "Economic Behavior in Institutional En- vironments: The Merger Wave of the 1980s." American Sociological Review 61 :699-718.

Stuart, Toby E., Ha Hoang, and Ralph C. Hybels. 1999. "Interorganizationa1 Endorsements and the Performance of Entrepreneurial Values." Administrative Science Quarterly 44:3 15-49.

Uzzi, Brian D. 1996. "The Sources and Conse- Test for Heteroskedasticity." Econometrica quences of Embeddedness for the Economic 48:817-38. Performance of Organizations: The Network Williamson, Oliver E. 1975. Markets and Hier- Effect." American Sociological Review archies. New York: Free Press. 61:674-98. Winship, Christopher and Robert D. Mare. 1992.

White, Halbert. 1980. "A Heteroskedasticity-"Models for Sample Selection Bias." Annual Consistent Covariance Matrix and a Direct Review of Sociology 18:327-50.

  • Recommend Us