Quantifying equilibrium network externalities in the ACH banking industry.

We seek to determine the causes and magnitudes of network externalities for the automated clearing house (ACH) electronic payments system. We construct an equilibrium model of customer and bank adoption of ACH. We structurally estimate the parameters of the model using an indirect inference procedure and panel data. The parameters are identified from exogenous variation in the adoption decisions of banks based outside the network and other factors. We find that most of the impediment to ACH adoption is from large customer fixed costs of adoption. Policies to provide moderate subsidies to customers and larger subsidies to banks for ACH adoption could increase welfare significantly.

1. Introduction

* Our goal is to estimate the size and importance of network externalities for the automated clearing house (ACH) banking industry using an equilibrium model of ACH usage and adoption. ACH is an electronic payment mechanism developed by the U.S. Federal Reserve and used by banks and customers. It is essentially an electronic alternative to paper checks, typically used for recurring transactions such as direct deposit paychecks and automated utility bill payments. Since banks on both sides of a transaction must adopt ACH for an ACH transaction to occur, ACH is a two-sided market.

As a two-sided market, ACH is characterized by network effects. A bank will be more likely to adopt ACH as other banks adopt because it will be able to originate more transactions with ACH, thus justifying the fixed costs of adoption. The importance of the network effect depends on the incremental profits to the bank from each ACH transaction relative to the fixed cost of adoption. Because banks most likely cannot compensate each other for ACH adoption, these network effects are network externalities and typically cause underutilization of the network good. The underutilization is particularly relevant for ACH--in an age when computers and technology have become prevalent, many more payments are performed with checks than with ACH.

Network effects may also exist at the level of the banks' ACH customers, which include employers, utility companies, and small businesses. These customers must also bear a fixed cost of adoption (e.g., updating their disbursement systems) and similarly may be more likely to adopt ACH if more of the customers with whom they transact also adopt. In contrast to banks, while it is clear that customers must adopt ACH to originate transactions, the extent to which a customer receiving an ACH transaction must actively adopt ACH is unknown. Starting direct deposit payroll, for example, takes very little effort on the part of the employee, the recipient in this case. This extent to which receiving customers must adopt, along with the magnitudes of customer and bank fixed costs of adoption and the customer and bank incremental benefits from ACH transactions, will all affect equilibrium adoption and may play a role in the low observed rates of adoption.

To understand the causes and extent of the network externalities, we specify a simple static two-sided market model of ACH technology adoption in local markets. Each market contains a set of banks, each with a given set of customers. Each customer must make a fixed number of transactions to other customers using either checks or ACH. While all banks and customers accept checks, some may not have adopted ACH. Some banks are locally based, while others are branches of big banks based outside the network. Local banks decide whether to adopt ACH based on whether the variable profits from ACH transactions conditional on adoption are greater than the fixed costs of adoption; the decisions of nonlocal banks are made exogenously and are known to the local banks. Following bank adoption, customers at banks that have adopted ACH choose whether or not to adopt ACH. We model two types of transactions: one-way transactions, which are passively accepted by the receiving customers, and two-way transactions, which require the receiving customers to adopt ACH actively.

The implications of the model depend on parameters that specify bank and customer costs and benefits, the proportion of two-way transactions, and other details including scales and time trends. We structurally estimate these parameters by applying an indirect inference estimator (a variant of the method of simulated moments) to bank-level panel data on ACH adoption and the number of ACH transactions. The idea of the estimator is to simulate data from the model (which requires solving for the equilibrium of the model conditional on structural parameters and unobservables) to find the parameters for which the simulated data most closely match the observed data.

This work builds on a recent literature on empirically estimating the extent of network effects for different industries (see Goolsbee and Klenow, 2002; Gowrisankaran and Stavins, 2004; Ohashi, 2003; Park, 2003; and Rysman, 2004a, 2004b, among others). Network externalities imply an interdependence of preferences, leading to simultaneity in equilibrium adoption decisions. This makes identification of the network externalities potentially difficult. These articles have all tried to find evidence of network effects by estimating correlations among usage decisions or by estimating reaction functions with techniques such as instrumental variables. Our work differs from this literature in that we fully specify an equilibrium model of interactions among banks and their customers, and we estimate the structural parameters of our model by computing and matching equilibrium predictions of the model to data. (1)

Our use of a fully specified structural model has a number of advantages. Relative to examining correlation among usage decisions (e.g., Gowrisankaran and Stavins, 2004), our structural model also allows us to estimate the magnitudes of the network effects, rather than just being able to test for their presence. In addition, we are able to allow for time-varying local shocks that otherwise might be incorrectly interpreted as network effects. Also, the methods that we develop here are novel and contribute to the literature on structural estimation of network games with potential multiple equilibria.

In comparison to articles that structurally estimate reaction functions using linear or log linear specifications (e.g., Rysman, 2004a), our fully specified equilibrium model has a number of advantages. First, it allows us to specify a strategic model of interactions with simultaneous decisions that is realistic given the discrete nature of adoption decisions in our industry. Second, our model allows us to understand the sources of the network effects. This is crucial for ACH policy analysis--for example, subsidies will have very different impacts on ACH usage depending on whether fixed costs are primarily at the bank level or at the customer level. Last, our structural model also allows us to perform a wider set of policy experiments, since it predicts the outcomes that will result from any policy intervention. This is particularly important to the extent that there are multiple equilibria: in this case, reaction functions are consistent with more than one outcome, implying that estimated reaction functions are not sufficient to simulate counterfactual equilibria.

Since it uses the same data as the present study, the work of Gowrisankaran and Stavins (2004) deserves particular attention. As mentioned above, that study takes a reduced-form approach to examining network externalities, modelling ACH adoption as a simple function of the percent of other banks that had adopted ACH in a local area. In contrast to the present article, the central point of the earlier work was to confirm the presence of network externalities statistically, not to measure the magnitude and sources (i.e., banks or customers) of the externalities. (2) Most relevant for the present article is that Gowrisankaran and Stavins (2004) proposed a number of different tests that would identify network effects separately from confounding factors. Although our methods and questions are different, this study builds on theirs in that our identification of the network effects is from some of the same sources. In addition, the earlier article guides us in our choice of indirect-inference moments--our indirect-inference estimator matches the coefficients from regressions that are similar to theirs evaluated on simulated equilibrium data. As an example, Gowrisankaran and Stavins (2004) propose a regression to exploit the quasi-experimental variation from the adoption decisions of small, remote branches of banks. Our analysis models this variation with more structure, allowing for exogenous nonlocal bank decisions and endogenous adoption decisions for their customers. We then choose structural parameters that match (among other things) the coefficients from regressions on real data that are similar to regressions from Gowrisankaran and Stavins (2004) to coefficients from the same regressions on simulated equilibrium data.

The remainder of this article is organized as follows. Section 2 describes the model. Section 3 describes the data. Section 4 details the estimation procedure, including the identification of the parameters. Section 5 provides results including policy experiments, and Section 6 concludes.

2. Model