The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Economic Review Federal Reserve Bank of San Francisco 1995 Number 1 Chan G. Huh and Bharat Trehan Modeling the Time-Series Behavior of the Aggregate Wage Rate Kenneth Kasa Comovements among National Stock Markets Elizabeth S. Ladennan Changes in the Structure of Urban Banking Markets in the West Table o f Contents Modeling the Time-Series Behavior of the Aggregate Wage Rate ..................... . 3 Chan G. Huh and Bharat Trehan Comovements among National Stock Markets ............................ ...................... . 14 Kenneth Kasa Changes in the Structure of Urban Banking Markets in the W est.............................. 21 Elizabeth S. Laderman Opinions expressed in the Economic Review do not neces sarily reflect the views of the management of the Federal Reserve Bank of San Francisco, or of the Board of Governors of the Federal Reserve System. The Federal Reserve Bank of San Francisco’s Economic Review is published three times a year by the Bank’s Research Department under the supervision of Jack H. Beebe, Senior Vice President and Director of Research. The publication is edited by Judith Goff. Design, production, and distribution are handled by the Public Information Department, with the assistance of Karen Flamme and William Rosenthal. For free copies of this and other Federal Reserve publicatons, write or phone the Public Information Department, Federal Reserve Bank of San Francisco, PO. Box 7702, San Francisco, California 94120. Phone (415) 974-2163. Printed on recycled paper with soybean inks. Modeling the Time-Series Behavior of the Aggregate Wage Rate This paper studies the behavior of wages relative to prices and productivity in a framework that places relatively few restrictions on the interactions among these va..-iables. The Chan G. Huh and Bharat Trehan Economist and Research Officer, respectively, Federal Reserve Bank of San Francisco. We would like to thank participants at a Federal Reserve System conference in Chicago, Tom Sargent, and editorial committee members for helpful comments on an earlier draft of this paper. Research assistance by Robert Ingenito is gratefully acknowledged. This paper looks at the time-series behavior of the real wage relative to that ofproductivity. Given an exogenous, nonstationary process for productivity, we use a simple model ofdynamic labor demand to show that the real wage and the marginal product of labor will be cointegrated if the representative firm chooses the profit-maximizing level of employment. Data for the postwar period satisfy this condition. On the basis of this result we estimate a vector error correction model containing prices, wages, andproductivity and examine the dynamic relationships among these variables. This specification provides a natural setting for looking at a number of issues of interest, including the role of the unemployment rate in the wage rate equation, issues of wage-price causality, and the effect of exogenous wage rate changes on productivity. use of the level of productivity as an anchor for the level of the real wage is a significant element of our analysis; it allows us to examine the behavior of wages without resorting to the common practice of arbitrarily detrending this variable and focusing on the residual. We use this specification to examine a number of key issues in the literature, including such questions as the relationship between wages and prices, the relationship between the unemployment rate and prices, and the way in which productivity shocks affect the wage rate. We begin with a condition that can be found in any simple model of competitive firm behavior-specifically, the firm sets the real wage equal to the marginal product of labor. In a time-series context this suggests that real wages and the marginal product of labor should move together over time. Starting with the assumption that the exogenous productivity process is nonstationary, we use a simple model of dynamic labor demand to showthat the real wage and the marginal product of labor will be cointegrated. Empirical tests over the post-war period reveal that the data are consistent with this condition. We go on to show that there exists a single cointegrating relationship among the nominal wage, the price level, and labor productivity. Cointegration allows us to cast the relationship among these vm:iables as a vector error correction model (VECM) and provides a natural framework for looking at a number of hypotheses about the wage rate. One issue has to do with the role of the unemployment rate in wage equations. We show that the role of the unemployment rate in such equations is sensitive both to the dynamic specification (in particular to whether the error correction term is included) and to the inclusion of a contemporaneous measure of productivity. Another issue has to do with the causal relationship between wages and prices. We find that prices Granger cause wages but that wages do not Granger cause prices. This is evidence against models that specify prices as a markup on wages. There does exist a non-negligible contemporaneous correlation among the innovations to these variables, and wage innovations could have an impact on prices if firms were to complete their adjustment to a wage shock within the quarter in which the shock occurred. 4 FRBSF ECONOMIC REVIEW 1995, NUMBER 1 Complete adjustment at such a rapid rate would appear to be at odds, however, with the large body of work on the sluggish behavior of prices. The remainder of the paper is organized as follows. Section I focuses on the long-run relationship between wages and productivity. We begin by laying out a simple model that allows us to derive testable restrictions upon the evolution of the real wage and labor productivity. Turning to the data, we first establish the univariate properties of the series and then examine the joint behavior of the real wage and productivity, with a view to determining whether the representative firm can be said to be on its long-run demand curve. Section II presents the estimated VECM; three sub-sections use this model to analyze the issues raised above as well as to examine the dynamic interrelationships among wages, prices, and productivity more generally. Section III concludes. I. THE RELATIONSHIP BETWEEN WAGES AND PRODUCTIVITY A Simple Model In this section we use a simple model to determine the kinds ofrestrictions that can be placed upon the behavior of the wage rate. We begin by assuming that the labor market is perfectly competitive. The model we employ is a version of the kind of dynamic labor demand model found, for instance, in Sargent (1978) or Nickell (1986). The representative firm produces a single output (whose price is taken as a numeraire) and maximizes the objective function given by 00 V = Et i~O WV t + j , where E t denotes the date t expectation and 13 is the discount factor for the firm. The one period profit function v t is (1) Vt = g(Lt, at) - W,L, - h (IlL t). g(.) denotes the production function in which labor (L) is the only input. We assume thatiJglaLt=gL>O andgLL<O. Labor is augmented by the technology shock term at, which can be thought of as measuring increases in knowledge that make the same unit of labor more productive over time. 1 We assume that at is nonstationary with positive drift, an assumption that is consistent with the empirical results below; W t denotes the real wage rate defined as 1. An alternative interpretation of this variable is that it represents the productive contribution of capital stock in the economy. W/P t' where Wt is the nominal wage and P t is the price of the firm's output; Il denotes the first difference, so that the function h(·) measures the cost of changing employment, which is borne by the firm. We assume the function h(·) to be such that there are symmetrical costs to both hiring and firing. The competitive solution can most easily be characterized as a situation in which the firm is on its dynamic labor demand curve Ld, which is implicitly defined by condition (2). The firm chooses L to maximize V, taking stochastic processes {at, w,} as given, so that (2) we == gL - [hL(IlL t ) - I3hL(E tIlLt+l)] , where the subscript c denotes the real wage rate (in terms of the product price) determined in a competitive labor market, hL = ah/aL, and the term in parentheses denotes the adjustment costs of changing the labor input in period t as well as in t+ 1. For our purposes, it suffices to note that, in equilibrium, (2) has to hold in each period. Equation (2) simply states that over time the marginal product of labor and the wage rate will move together, apart from deviations caused by the costs of changing employment. The nature of the adjustment costs will determine the relative behavior of the wage rate and productivity over time. By the symmetry assumption, both an increase and decrease in employment incur labor adjustment costs, which in turn depend only on the net amount of change (IlL t). Thus, the term in parentheses in (2) above is likely to be small and temporary for both constant or trending levels of employment, because it is the difference between the marginal cost of adjustment in two adjacent periods. Despite the apparent intuitive appeal of condition (2), there are alternative theories of the wage-employment determination mechanism that cannot be characterized in this way. The efficient wage bargaining theory that focuses on unionized labor markets is a case in point. Both the static version of the efficient wage bargaining model (e.g. , Abowd 1987, MaCurdy and Pencavel 1986), and its dynamic extension (Espinosa and Rhee 1989), emphasize and focus on the strategic nature of the interaction between firms and workers in determining wages and employment. In these models, factors such as the relative strength of unions and firms and the union's preferences over wages versus employment are crucial determinants of the actual bargaining equilibrium outcome. There is a critical implication of this view that is relevant to our exercise. Typically, the solution set of wage-employment pairs for either the static or the dynamic efficient bargaining problem does not include points that lie on the representative firm's demand-for-Iabor curve. Thus, (2) HUH AND TREHAN/TIME SERIES BEHAVIOR OF THE AGGREGATE WAGE RATE does not apply to this situation. 2 The fact that unions have played a significant role in the U. S. over our sample period provides a priori grounds for considering the efficient bargaining model seriously. In other words, there appears to be sufficient reason to believe that the behavioral prediction of (2) might not hold for our sample period. TABLE 1 UNIT ROOT TESTS AGAINST ALTERNATIVE OF While our discussion above has been in terms of the marginal product of labor, we have data only on average product. How closely does the latter approximate the former? The answer depends upon the underlying production function. If the production function is Cobb-Douglas, for instance, the log of the marginal product is just a constant plus the log of the average product. More generally, the results below will go through if there is a linear relationship between the log of the average and marginal products of labor. We can relate the two measures by introducing the elasticity of output with respect to labor, which is defined as the ratio between the marginal product and the average product. Thus, the log of the marginal product of labor equals the log of the average product plus the log of the elasticity, which is a constant for a wide range of production. function specifications (e. g., linear and CES). The measure of average product we use is the output per hour of all persons in the business sector, which is compiled by the Department of Labor. The log of this series is denoted by LYHR. For wages we use compensation per hour in the business sector, and denote the log of this variable by LNWAG. This measure seems most relevant to our purposes, since it includes total payments made by firms to all workers. Our focus upon labor productivity as an anchor for the real wage implies that we need a measure of the product wage, that is, a real wage measured in units of the firm's output. Consequently, we use the implicit price deflator for business sector output (whose log we denote by LDEF) to deflate the nominal wage. We use LRWAG to denote the log of the real wage. All series are available on the Citibase data tape. The first order of business is to establish some facts about the long-run behavior of the individual series. Accordingly, Table 1 presents tests of the unit root hypothesis for each of 2. Among models of unionized labor markets, the monopoly union model provides an exception. According to the model, the monopoly union unilaterally chooses the wage rate, leaving the employment decision entirely up to the firm; the firm, in tum, takes the wage rate as given and determines employment using its demand-for-labor curve. However, contrary to this model, in practice bargaining usually takes No TREND BREAK A. loG LEVELS Variable A First Look at the Data 5 LYHR LNWAG LDEF LRWAG Dickey-Fuller Test -1.74 -1.27 -1.43 -1.12 Phillips -1.19 -1.84 -1.88 -1.07 B. DIFFERENCES Variable dLYHR dLNWAG dLDEF dLRWAG Dickey-Fuller Test Phillips -4.94** -3.19* -3.26* -4.43** -12.76** -11.56** -9.33** -12.74** NOTES: Regressions in panel A contain a constant and a time trend. For both test statistics reported here, the 10 percent significance level is - 3.15. Regressions in panel B contain a constant only; * denotes significance at 5 percent; ** denotes significance at 1 percent. Dickey-Fuller test equations contain four lags of first difference of the dependent variable. To compute the Phillips test statistics, we use Schwert's (1987) /12 formula, which implies the use of thirteen autocovariances. these series against the alternative that it can be described as stationary around a linear trend over the 1948.Q2-1990. Q3 sample period. The results for the Dickey-Fuller test reveal that in no case are we able to reject the unit root hypothesis at even the 10 percent significance level. 3 Table 1also contains results from Phillips' (1987) test. This test al1lows the error term to follow a more general process than the DickeyFuller test. It turns out that the second test leads to the same results as the first. While we have allowed for a linear trend in these tests, visual inspection of the data suggests that the trend growth place over both wages as well as employment. Further, Espinosa and Rhee (1989) show that the outcome described by the monopoly union model is unlikely to occur in a general dynamic bargaining game setup. There is a set ofPareto superior solutions that dominate such an outcome in a repeated game context. 3. In the case of the Dickey-Fuller test, the test statistic is calculated as the ratio ofthe coefficient of the lagged level to its standard error (from a regression where the first difference is regressed on a constant, a time trend, lagged first differences, and a lagged level); critical values are available in Fuller (1976). 6 FRBSF ECONOMIC REvIEW 1995, NUMBER 1 rate of LYHR and LRWAG may have changed over the sample. Indeed, the productivity slowdown over this period has been widely noted. Consequently, we test whether the unit root specification (with no change in drift) can be rejected against an alternative that allows for a single change in a deterministic trend. This specification has been suggested as a reasonable alternative (in a somewhat different context) by both Rappoport and Reichlin (1989) and Perron (1989). Here we implement a procedure suggested by Christiano (1988). Specifically, we employ his "min-ta " procedure. 4 The results from this procedure are shown in Table 2. For LYHR, for example, the procedure finds the most likely break date to be 1964.Q1, where the computed t statistic has a value of - 3.45. However, this value is considerably smaller than the expected value obtained under the null, and the computed t statistic has a marginal significance level of .79. Similar results are obtained for the other variables. Thus, in no case are we even close to rejecting the unit root null against the alternative of a break in a deterministic trend. Are Firms on Their Demand Curves? The results in Tables 1and 2 imply that the individual series contain unit roots. What can we say about the joint behavior of these series? Equation (2) implies that gL (labor's marginal product) and We (the real wage in a competitive labor market) will be cointegrated. This result 4. The intuition behind the procedure is as follows. In attempting to determine whether a break has occurred, the date of any potential break is usually determined after looking at the same data. However, using a (formal or informal) search procedure to determine the break date implies that the distribution of the resulting test statistic will no longer be the same as it would be if the break date had been determined independently of the data at hand. Christiano suggests a number of alternative techniques to choose the most likely date for a break in the trend and then constructs empirical distributions which take this "pretesting" into account. The null hypothesis is that the process in question contains a unit root, while the alternative is AYt = ao + aPT + bo*trnd + blDT*trnd + QYt-1 + d(L) AYt-1 + Vt O<t<T where = 1, T';; t.;; T, and trnd denotes a linear time trend. Thus, this specification allows for both a jump in the level ofthe variable and a change in the slope at date T. The value of T is then allowed to vary over the entire sample (that is, we allow each date in the sample to be the break date), and we compute the value of tot at each date. We then define the date at which ta attains its minimum value as the most likely break date. An empirical distribution for this statistic is obtained by using the bootstrap to construct 1000 new series. TABLE 2 UNIT ROOT TESTS AGAINST ALTERNATIVES THAT ALLOW FOR BREAK IN TREND SIGNIFICANCE VARIABLES LYHR LNWAG LDEF LRWAG MOST LIKELY LEVEL/EXPECTED BREAK DATE VALUE 1964.Ql 1957.Q2 1959.Q4 1969.Q2 -3.45 -2.01 -2.81 -3.45 .79/ -4.02 1.0/ -4.00 1.0/ -4.02 .80/-4.00 NOTES: All regressions allow both the constant and time trend to change over the sample. Calculation of the Most Likely Break Date excludes three years of data at either end of the sample. The dates reported are those that lead to the smallest t statistic on the lagged level of the dependent variable. is intuitive; it says that ifthe firm stays on its demand curve, the marginal product of labor and the wage rate should move together over time, apart from temporary deviations caused by adjustment costs. It should be noted that temporary deviations between labor productivity and the wage rate can occur for a number of reasons besides the costs of adjusting employment. For example, Bils (1990) shows that in sectors with long-term contracts, real wages increase significantlyrelative to wages elsewhere in the economy-in the first year of the contract, but then decline over the life of the contract. Our empirical approximation of the marginal product of labor by the average product is another reason for not expecting an exact relationship between LYHR and LRWAG, even if the firm is on its demand-for-labor curve in the long run. Turning to the data, a regression of the real wage on labor productivity over the 1947.Q1-1990.Q3 period leads to LRWAG t = -0.01 + 0.995 LYHR,. The Augmented Dickey-Fuller (ADF) test (see Engle and Granger 1987) leads to a test statistic of - 5.02, which compares to a 1 percent significance level of - 3.73. 5 Use of Johansen's maximum likelihood procedure leads to the same result. Under the null hypothesis that there is no cointegrating vector, the computed value of the trace test is 5. The critical values are from Engle and Yoo (1988). The equation used to estimate the test statistic contains one lag of the dependent variable. The lag length was arrived at by starting with six lags and eliminating HUH AND TREHAN /TIME SERIES BEHAVIOR OF THE AGGREGATE WAGE RATE 39.2, which is significant at the 1 percent level. The statistic for the maximal eigenvalue test (with a computed value of 31.0) also is significant at 1 percent. 6 (See Johansen and Juse1ius 1990 for a discussion of the tests and tabulated critical values.) These results reveal that labor productivity and the observed real wage have shared the same stochastic trend component during our sample period and consequently are consistent with models that imply that the representative firm is on its demand-for-Iabor curve. The test for cointegration between real wages and productivity is actually a test for cointegration among nominal wages, prices, and productivity, where the coefficient on (the log of) the price level is set equal but opposite in sign to the coefficient on wages. An alternative way to proceed is to estimate the cointegrating vector among these three variables without imposing any restrictions. Removing this restriction introduces prices explicitly into the model and allows us to look at a larger system in the analysis that follows. One advantage of looking at this system is that we can now allow for the possibility of exogenous shocks to the price level. For example, we can now allow a union's actions to affect LNWAG, while LDEF is influenced by the actions of the monetary authority or by OPEC. More generally, the point is that we can now allow for a greater number of disturbances to the model. Repeating the cointegration test above leads to LNWAGt = -4.56 + 1.02 LDEFt + 0.97 LYHRr We obtain a test statistic of - 5.02, which is again significant at 1 percent. 7 Increasing the number of variables in the model raises the possibility of more than one cointegrating relationship among these variables. We use Johansen's maximum likelihood approach to determine the number of cointegrating vectors. Under the null hypothesis that no cointegrating relationship exists among these variables, the value of Johansen's trace test is 48.5, which is significant at 1 percent. 8 By contrast, we cannot reject the null hypothesis that there is at most one cointegrating vector at the 10 percent level. (The computed value of the trace-test statistic is 15.3.) The maximal eigenvalue test the insignificant terms. (In no case do we fail to reject the null of no cointegration. ) 6. The estimated model contains six lags of the first-differenced variables; the error terms from these equations are well-behaved. 7. The residual equation contains one lag. 8. The model included six lags as before. However, the residuals indicate that the normality assumption is violated because of too much kurtosis. Induding up to four more lags does not solve this problem. Gonzalo (1989) points out that the Johansen procedure is robust to violations of the normality assumption. 7 yields the same results. These results indicate that there are two distinct stochastic trends driving these three variables, and that there is one cointegrating vector. With the variables ordered as LNWAGt , LDEFt , and LYHRt , the estimated cointegrating vector is (1 - .995 - .951). Recent research has shown that the estimates obtained from the Johansen procedure are superior to those obtained from the OLS regression; consequently, this is the cointegrating vector that \ve "viII employ belov/. 9 II. THE ESTIMATED VECM AND ITS ApPLICATIONS The finding that the real wage rate and productivity are cointegrated implies that we can specify the model as a vector error correction model (VECM).lO The equations of the estimated VECM are shown in Table 3. Notice that the error correction term (EC t -1) enters significantly in the LNWAG equation and has a negative sign. This implies, for example, that the nominal wage falls whenever it gets too high relative to the price level and the productivity of labor. By contrast, ECt _ 1 does not enter significantly into either the LDEF or the LYHR equation; it actually has the wrong sign in the LYHR equation. This implies that, in general, it is LNWAG that adjusts to correct the "error" among these variables. We will return to this issue below. Note also that the adjusted R2 of the ~LYHR equation is close to zero, implying that the model does not do a very good job of explaining changes in productivity. In the rest of this section we use this model to study several issues regarding the behavior of wages as well as the interaction of wages, prices, and productivity. The first extension is to introduce the unemployment rate into the model and to examine the relationship between unemployment and wages. This allows us to look at the "Phillips curve" in a framework that does not impose a potentially artificial separation between the short and long run on the data. Note that the inclusion of the unemployment rate in the equation for wages also can be motivated by appealing to efficiency wage theories of wage determination. Next, we look at the relationship between wages and prices. Recent papers by Gordon (1988) and Mehra (1991) have looked at the causal relationships between these variables. Our model offers an alternative way of examining these issues. Finally, we will use the model to examine the dynamic relationships among wages, prices, and productivity. 9. See Gonzalo (1989), for example. 10. See Granger's Representation Theorem (Engle and Granger 1987). 8 FRBSF ECONOMIC REvIEW 1995, NUMBER 1 TABLE 3 Introducing the Unemployment Rate THE ESTIMATED VECM Perhaps the most straightforward way of introducing the unemployment rate in our wage equation is by appealing to Phillips (1958), who modeled the change in wages as a function of the unemployment rate. A recent example of an analysis carried out along these lines is Vroman and Abowd (1988), who regress the growth in hourly earnings on alternative measures of the civilian unemployment rate, the lagged consumer price index, and some other variables. The literature on efficiency wages provides another motivation. In Shapiro and Stiglitz (1984), for example, firms are unable to monitor workers' efforts fully. To offset this and provide the right incentive to workers, a firm has to pay a wage rate that meets a "no shirking constraint." The prevailing unemployment rate is a key determinant of such a wage rate and is inversely related to it, because the prevailing unemployment rate affects the probability of a worker finding a new job if he is found shirking and fired. 11 It is not difficult to motivate the inclusion of the unemployment rate in the price equation, either. For example, specifying prices as a cyclically sensitive markup on wages would imply a relationship between prices and the unemployment rate as well. Our strategy is simply to include the unemployment rate into the VECM presented earlier. 12 While the resulting specification will not include many of the wrinkles of recent Phillips curve analyses, it improves upon conventional specifications in two ways. First, it allows changes in productivity to affect wages and prices directly. Second, it allows us to examine the cyclical relationship between wages and the unemployment rate in a framework that also models the long-run behavior of these variables, rather than assuming it away either by linear detrending or by first differencing the data. Consider first what happens when six lags of the level of the unemployment rate are included in our model. The null that the (lagged) unemployment rate does not belong in the price equation can be rejected only at a marginal significance level of 60 percent, while the null that it does not EXPLANAIDRY VARIABLES IlLNWAG t Constant -0.54 ( -3.0) IlLNWAGt~l IlLNWAG t _ 2 IlLNWAGt~3 IlLNWAGt~4 IlLNWAG t _ 5 IlLNWAGt~6 IlLDEFt~l IlLDEFt~2 IlLDEFt _ 3 IlLDEFt~4 IlLDEFt~5 IlLDEFt _ 6 IlLYHRt~l IlLYHRt~2 IlLYHRt~3 IlLYHRt~4 IlLYHRt~5 IlLYHRt _ 6 ECt -l R2/ adj. R2 S.E.E. (x102) Q(36)/SIG. LEVEL IlLDEFt 0.19 (1.1) IlLYHRt -0.22 (1.0) 0.13 (1.3) 0.11 (1.2) -0.12 ( -1.2) 0.04 (0.4) 0.14 (1.5) 0.07 (0.8) -0.11 ( -1.3) 0.21 (2.5) 0.08 (0.9) 0.01 (0.1) 0.02 (0.2) ( -0.2) 0.15 (1.3) 0.04 (0.3) -0.12 ( -1.0) 0.10 (0.9) 0.01 (0.1) 0.08 (0.8) 0.37 (3.4) 0.22 (2.0) -0.05 ( -0.5) -0.31 ( -2.8) 0.05 (0.4) 0.06 (0.6) 0.46 (4.5) 0.19 (1.8) -0.02 ( -0.2) -0.06 ( -0.6) -0.08 ( -0.8) 0.10 (1.0) -0.24 ( -1.8) 0.05 (0.4) -0.04 ( -0.3) -0.22 ( -1.6) -0.07 ( -5.0) -0.06 ( -0.5) -0.14 (-1. 7) 0.03 (0.3) -0.11 ( -1.3) -0.19 ( -2.4) -0.13 ( -1.7) 0.09 (1.1) 0.15 (1.9) 0.01 (0.1) -0.06 ( -0.8) 0.03 (0.5) -0.17 ( -2.2) 0.01 (0.1) -0.10 ( -1.0) 0.05 (0.5) 0.02 (0.2) -0.31 (-3.1) 0.02 (0.2) -0.06 ( -0.6) -0.12 ( -3.0) 0.04 (1.1) -0.05 ( -1.0) -0.D2 .41/.33 .49/ .42 .16/ .06 0.70 0.65 0.87 32.7/.63 9.1/.79 6.8/.87 NOTES: t statistics are in parentheses. The error correction term is ECt = LNWAGt - .995 LDEFt - .951 LYHRt · 11. Also see Blanchard and Fisher (1989) pp. 455-463 for a more general discussion of efficiency wage theories. 12. We began by considering the time-series properties ofthe unemployment rate. We carried out the two tests in Table 1 in order to test for stationarity. While it is possible to reject the null that the unemployment rate contains a unit root (at the 6 percent level of significance) when the augmented Dickey-Fuller test is used, we fail to reject when the Phillips test is used. Given this conflict, we chose to go with our prior, which is that the unemployment rate is stationary. Accordingly, we decided to include the level, and not the first difference, of the unemployment rate in the VECM. Another alternative would be to model the unemployment rate as being stationary around a shifted mean. See Evans (1989). HUH AND TREHAN /TIME SERIES BEHAVIOR OF THE AGGREGATE WAGE RATE belong in the wage equation can be rejected only at a marginal significance level of 79 percent. The error correction term remains significant in the wage equation (with its estimated coefficient getting noticeably larger in absolute terms) and insignificant in the price equation even after the unemployment rate is introduced. As an alternative, we tried including the log levels of the unemployment rate in the VECM. Once again, the null that the unemployment rate does not belong in the equation cannot be rejected in either case at the 50 percent level of significance. Thus, we find no evidence to suggest that the lagged unemployment rate should be included in either of these equations. The experiments described so far do not match up precisely with the Phillips curve literature, since we have omitted the contemporaneous unemployment rate. Note that introducing the contemporaneous unemployment rate into this systemis not innocuous, since it begs the question of whether the unemployment rate is exogenous. However, we decided to include the contemporaneous term in order to allow a direct comparison with the Phillips curve literature. An F test on the contemporaneous and six lagged values of the unemployment rate fails to reject the null that these terms are zero at the 35 percent level of significance. However, the contemporaneous unemployment rate term has a t statistic of - 2. I in the wage equation (and -1.8 ill the price equation). In addition to the problem of exogeneity discussed above, another problem in trying to gauge the significance of this result is that the unemployment rate is the only contemporaneous variable included in the wage equation. Thus, its importance may result from the fact that it is the only way that contemporaneous developments are allowed to affect wages. There is an easy way around this problem in our model: Specifically, we introduce the contemporaneous change in productivity into the wage equation and see what effect this has on the significance of the unemployment rate. 13 It turns out that doing so reduces the t statistic on the contemporaneous unemployment rate to - 1. 5, and we cannot reject the null that both contemporaneous and lagged unemployment terms are zero at the 70 percent level of significance. By contrast, the contemporaneous productivity term has a t statistic that is close to 3. It also is worth mentioning that the error correction term remains significant in the wage rate equation through all the exercises described above. Finally, if we drop the error correction term from the specification just described, the t statistic on the contemporaneous unemployment rate goes to - 2 while the null that the current and lagged unemployment rate terms are zero can be rejected at 11 percent. 13. We consider the issue of whether LYHR is predetermined with respect to LNWAG below. 9 Our results demonstrate that inferences regarding the inclusion of the unemployment rate in an equation for wages are sensitive to how the dynamics of the wage rate are specified, as well as to whether the contemporaneous effects of changes in productivity are taken into account. While our search has not been exhaustive, we have shown that the unemployment rate is not very important in explaining the wage rate in a framework where the long-run behavior of wages is modeled explicitly. However, we do not wish to claim that unemployment can never matter for wages within our framework. Instead, we prefer to thillk of this exercise as an illustration of the usefulness of studying a "cyclical" relationship (between wages and the unemployment rate in this case) in the context of a model that ties down long-run behavior (here, of the wage rate). Wage-Price Causality Our model also provides a straightforward way to study another set of issues, namely, the relationship between wages and prices. Causal relationships between these variables can be motivated in a number of ways. For example, Keynesian models commonly specify prices as a markup over wages. In these models a permanent change in the level of wages will have a permanent effect on the level of prices. Similarly, it is not hard to find models of the real wage rate in which nominal wages react to price innovations. Two papers that recently looked at the empirical relationship between wages and prices are Gordon (1988) and Mehra (1991). Gordon looks at the relationship between prices and unit labor costs, with the latter variable defined as the difference between nominal wages and an exogenous (piecewise linear) trend in productivity. He concludes that wages and prices are determined independently of each other, though the evidence that wages do not have much effect on prices is stronger than the other way around. More recently, Mehra (1991) has carried out a similar analysis. In his model, prices are specified as markups over productivity-adjusted labor costs and are subject to various shocks. Wages are specified as a function of cyclical demand and expected prices. He then goes on to discuss how such equations imply that wages and prices must be related in the long run. Mehra carefully analyzes the time-series behavior of individual series and finds that the two are integrated of order 2, and that it is the first difference of wages that is cointegrated with the first difference of prices. He finds that the rate of inflation is Granger causally prior to the rate of change of wages, not vice versa. The VECM specification we employ here allows us to look at the long-run relationship between wages and prices as well. In addition, our specification allows a potential 10 FRBSF ECONOMIC REvIEW 1995, NUMBER 1 role for feedback from wage or price shocks to productivity (an issue we will return to below). Before going further, it is worth pointing out that the error correction term we employ has an interesting antecedent in a term used in Gordon (1988). Specifically, Gordon includes the difference between the lagged level of trend unit labor costs and the price level in equations for both the rate of change of prices and of trend unit labor costs, and he interprets this term as labor's income share. It turns out that t.~is term does not enter significantly into either the inflation or unit labor cost equations. We begin by asking about the nature of the long-run adjustments between these variables. First, does either of these variables adjust to maintain the long-run relationship estimated above? Table 3 shows that the estimated error correction term does not enter significantly into the price level equation. We obtain the same result when we use the test discussed in Johansen and Juselius (1990); the X2 statistic calculated under the null that the error correction term does not belong in the price equation has a marginal significance level of .3. By contrast, restricting the error correction term to be zero in the wage equation leads to a X2 (1) statistic of 19, which is significant at any reasonable level. Thus, it is the level of wages-and not the price level-that adjusts to maintain the cointegrating relationship in our model. This influence of prices on wages through the error correction term means that the common practice of estimating a single equation where the price level is regressed on the contemporaneous wage rate (and other variables) is inappropriate. (See Banerjee, et al., 1993 for a discussion ofthe issues involved.) The appropriate way to proceed in studying this issue would be to estimate this equation as part of a system that also includes an equation for the wage rate. It is, of course, still possible that changes in the growth rate of wages temporarily affect the growth rate of prices. However, the data do not provide much support for this hypothesis either (the computed F statistic has a marginal significance level of 18 percent). By contrast, we cannot reject the hypothesis that changes in inflation affect the growth rate of wages (we obtain an F statistic with a marginal significance level of 1 percent). Thus, we find that prices Granger cause wages but that wages do not Granger cause prices. Our evidence against the wage markup model echoes the results of both Gordon and Mehra, although the results are not exactly the same. 14 14. These differences probably reflect both differences in specification (including the precise variables used) as well as differences in modeling strl;Ltegy. Of the latter, it is worth noting that our specification does not include dummies or other exogenous variables. The Dynamics of the Wages-Productivity Relationship We now look at how wages, prices, and productivity respond to various disturbances to the system. Our VECM can be used to analyze a number of interesting issues, some of which are related to the issues raised above. For instance, we can use the model to estimate the responses to a permanent change in productivity. Do firms react to productivity shocks by raising the nominal wage, or is the resulting long-run increase in real wages achieved by falling prices? Similarly, do shocks to productivity have a significant effect on the real wage or are nominal wage innovations more important for the real wage? To study these and related questions we use the unrestricted VECM presented inTable 3 above. As a robustness check, we also looked at two simplified versions of our model. First, we used a statistical criterion to select lag lengths; however, the resulting shorter lag lengths did not lead to noticeable changes in the dynamic responses obtained from the model. Second, we also estimated a model that imposed the long-run restrictions that we tested for above; specificaHy, the model excluded the error correction term from the price and productivity equations. We will point out any difference in the dynamics below. There is still the matter of identification. While there are a number of alternative ways of imposing identifying restrictions on vector autoregressions, none is completely unproblematic; see Hansen and Sargent (1989) for a recent critique. Here, we present results using the earliest such method of identification, suggested by Sims (1980), with the hope that we can get at some of the issues we are interested in by using relatively simple restrictions. 15 Given our concern with productivity and the wage rate, we examined two orderings that alternatively place productivity and nominal wages at the top of the system. Our discussion below focuses on the case where productivity is placed first; however, we also discuss how the results differ in the case where wages are placed first. 16 Figure 1 shows impulse responses from the system where productivity is placed first, prices are placed second, and wages are placed last. The top left panel shows the response of LRWAG and LYHR to productivity shocks, 15. Given two variables X and Y, for example, one could leave the residuals from the equation for X unchanged and transform the residuals from the Y equation so that they are orthoganal to those from the X equation. Thus, the covariance between the estimated error terms is attributed to the innovation in X, and X is said to be ordered first. 16. The correlation between LDEF and LYHR residuals is - .33, that between LDEF and LNWAG residuals is .43, and that between LYHR and LNWAG residuals is .30. HUH AND TREHAN / TIME SERIES BEHAVIOR OF THE AGGREGATE WAGE RATE FIGURE 1 DYNAMIC RESPONSES (ORDERING: LYHR, LDEF, LNWAO) LYHR INNOVATIONS 240 I 40 [ LNWAG ~~~---=======~ 180 LYHR ....... ............. ' -80 120 .". '" '" LDEF .... ""."""""." ... .. . "" " -140 60 ..... -200 10 30 20 50 40 60 1.u.L ........... 10 20 30 40 Quarters 50 60 Quarters LDEF INNOVATIONS 270 o~----------------, -60 210 LRWAG .... "" .. """" . -120 150 -180 90 -240 I.u.a. ........... 10 30 20 40 50 60 30 .. , .' .. ' LDEF .. .. 10 ....... ................... 30 20 40 50 60 LNWAG INNOVATIONS 120.....----------------, 180.....-----------------, 120 60 .... LDEF LRWAG .., 60~ .... Ol-.~·' .••••• TVHR '. :-O--~L-N=W...A - : G - - - - - - •.......... ~~'.:.:.' . .~ •• ~ . . ':.:..:,' '.=.:." Ld -60 . ................... .... . -120 I.u.a. 1 . o IT·:...---------------; ........... 10 20 30 40 50 60 10 20 30 40 50 60 11 12 FRBSF EcONOMIC REVIEW 1995, NUMBER 1 while the top right panel shows the individual responses of LNWAG and LDEF. (The LRWAG response is obtained as the difference between the LNWAG and the LDEF responses.) The left panel shows that the effects of productivity innovations grow over time; the real wage changes little over the first four to six quarters but does catch up with the change in productivity after a while. The right panel reveals that nominal wages do not go up very much following a positive shock to productivity; instead, the required increase in the real wage is achieved by a fall in the price level. Such a response might occur, for instance, if firms tend to introduce improved products or technology at the old prices. In the alternative ordering, where wages are placed first, productivity second, and prices third, positive productivity shocks also lead to permanently higher real wages because of a reduction in the price level; however, anomalously, the nominal wage falls somewhat. The middle right panel shows that price level shocks are persistent as well, and that they tend to grow over time. It also shows that while the nominal wage does go up in response to the price shock, it never catches up. The outcome, shown in the middle left panel, is a permanently lower real wage (LRWAG). Thus, price level surprises are associated with lower productivity and real wages. A similar result is obtained with the alternative ordering. Such a response might result, for instance, if a negative supply shock manifests itself first as an increase in prices. The panels at the bottom show the effects of a positive nominal wage shock. While the nominal wage is persistently higher as a result, the price level increases by more than the increase in wages, so that the ultimate outcome is a reduction in both the real wage and productivity. We obtain a similar result in the system with the alternative ordering, although the nominal wage shock has a larger effect on LNWAG and LDEF than in the first ordering. As might be expected, the effects of the nominal wage shock are sensitive to whether the error correction term is included in the price and productivity equations, regardless of the ordering. If this term is not included, both the real and nominal wage return to zero over a six to seven year horizon following a nominal wage shock (in the LYHR, LDEF, LNWAG ordering). The initial response of the price level in that model also is much smaller than that shown in the bottom panels. The variance decompositions associated with Figure 1 are shown in Table 4. LYHR appears to be largely exogenous. Note also that in the long run, LNWAG is driven largely by LDEF innovations. LDEF is driven largely by its own innovations, with LYHR innovations playing a small role. The bottom panel shows that while nominal wage innovations have a substantial impact on real wages in the short run, they become less important as the time horizon TABLE 4 VARIANCE DECOMPOSITIONS ORDERING: LYHR LDEF LNWAG QUAlITERS AHEAD LYHR LDEF LNWAG LRWAG LYHR LDEF 4 8 12 60 100 99 92 87 66 26 1 4 8 12 60 10 5 9 12 16 90 94 88 84 77 1 4 8 12 60 9 2 1 0 0 81 86 94 1 4 8 12 60 34 36 54 71 66 10 7 12 10 25 1 0 1 7 11 31 72 LNWAG 0 0 1 2 8 0 1 3 4 7 60 23 19 14 t:- v 56 57 34 19 8 NOTE: This table reports the percentage of forecast error variance that is attributed to each of the three shocks. lengthens, while productivity innovations become more and more important. The alternative ordering does lead to a greater role for LNWAG innovations in both the LNWAG and LDEF forecasts. For instance, at a horizon of 60 quarters, wage innovations are somewhat more important than price innovations for predicting wages and are roughly as important for predicting prices. However, LNWAG innovations explain almost none of the forecast error variance of LYHR or LRWAG in the long run. And neither LDEF nor LNWAG innovations account for much of the variation in LYHR. Overall, despite differences between the two orderings, they share a number of features. Thus, in both orderings productivity shocks affect the real wage rate through price level adjustments; price level innovations lower the real wage, and nominal wage shocks have little effect on the real wage in the long run. Before concluding this section it also is worth reviewing some of the evidence presented here in light of the results HUH AND TREHAN / TIME SERIES BEHAVIOR OF THE AGGREGATE WAGE RATE 13 discussed in prior sections. Our results indicate that shocks to the price level have a significant effect on wages. By contrast, the proportion of the price level forecast error attributable to the nominal wage shock is relatively small, even when the nominal wage is placed first. There is, of course, significant contemporaneous correlation between the innovations to these two variables; if this correlation were assumed to be the result of shocks to the wage rate, then wage shocks couid be said to have a non-~egligible effect on pricesP However, this inference can be reconciled with the Granger causality tests presented above only if firms complete the required price adjustment (to wage shocks) within the quarter in which the wage shocks occur. Such rapid adjustment seems rather unlikely to us. Finally, there is no evidence to suggest that shocks to the nominal wage rate have any permanent effect on either real wages or productivity. REFERENCES ill. Espinosa, Maria Paz, and Changyong Rhee. 1989. "Efficient Wage Bargaining as a Repeated Game." Quarterly Journal ofEconomics (August) pp. 565-588. SUMMARY AND CONCLUSIONS We have argued that the time-series behavior of aggregate wages should be studied in relation to the time-series behavior of productivity. If productivity is nonstationary, relatively tight restrictions on the joint behavior of these variables can be derived from models in which the representative firm is on its demand curve in the long run. We find that data for the postwar U.S. economy are consistent with the hypothesis that the representative firm is on its demand-for-Iabor curve in the long run. This finding-which in terms of the empirics is that productivity, wages and prices are cointegrated-allows us to cast the data in the form of a vector error correction model. Using this specification we find that (Granger) causality runs from prices and productivity to wages but not the other way around. Further, our analysis reveals that the measured impact of cyclical variables, such as the unemployment rate, is sensitive both to how the long run is modeled and to the inclusion of a measure of productivity in the wage equation. Our analysis also reveals that nominal wage innovations have little, if any, influence on the long-run behavior of real wages (or, by implication, of productivity). Instead, the long-term behavior of the real wage rate is determined largely by innovations to productivity, and these innovations act almost entirely through changes in the price level. Abowd, 1.M. 1987. "Collective Bargaining and the Division of the Value of the Enterprise." NBER Working Paper No. 2137. Banerjee, Anindya, Juan Dolado, John W. Galbraith, and David E Hendly. 1993. Cointegration, Error-Correction, and the Econometric Analysis ofNon-stationary Data. New York: Oxford University Press. Bils, Mark. 1990. "Wage and Employment Patterns in Long-Term Contracts when Labor Is Quasi-Fixed." NBER Macroeconomics Annual, pp. 187-226. Blanchard, Olivier, and Stanley Fischer. 1989. Lectures on Macroeconomics. Cambridge: MIT Press. Christiano, Lawrence 1. 1988. "Searching for a Break in Real GNP." NBER Working Paper No. 2695. Engle, Robert E, and C. W. 1. Granger. 1987. "Cointegration and Error Correction: Representation, Estimation and Testing." Econometrica (March) pp. 251-276. Engle, Robert E, and Byung Sam Yoo. 1988. "Forecasting and Testing in Co-integrated Systems." Journal of Econometrics, pp. 143159. Evans, George W. 1989. "Output and Unemployment Dynamics in the United States:1950-1985." Journal ofAppliedEconometrics 3, pp. 213-237. Fuller, Wayne A. 1976. Introduction to Statistical Time Series. New York: John Wiley & Sons. Gonzalo, Jesus. 1989. "Comparison of Five Alternative Methods of Estimating Long-Run Equilibrium Relationships." Unpublished manuscript. University of California San Diego (November). Gordon, Robert 1. 1988. "The Role of Wages in the Infll(ltion Process." American Economic Review Papers and Proceedings, pp. 276283. Hansen, Lars P., and Thomas 1. Sargent. 1989. "Two Difficulties in Estimating Vector Autoregressions." In Rational Expectations Econometrics. Boulder: Westview Press. Johansen, Sl'lren, and Katarina Juselius. 1990. "Maximum Likelihood Estimation and Inference on Cointegration-with Applications to the Demand for Money." Oxford Bulletin of Economics and Statistics, pp. 169-210. MaCurdy, Thomas E., and John H. Pencavel. 1986. "Testing between Competing Models of Wage and Employment Determination in Unionized Markets." Journal ofPolitical Economy, pp. S3-S39. Mehra, Yash. 1991. "Wage Growth and the Inflation Process." American Economic Review, pp. 931-937. Nickell, Stephen. 1986. "Dynamic Models of Labor Demand." In Handbook ofLabor Economics, eds. Orley Ashenfelter and Richard Layard. Amsterdam: Elsevier Science Publishers B. V. Perron, Pierre. 1989. "The Great Crash, the Oil Price Shock and the Unit Root Hypothesis." Econometrica, pp. 1361-1401. 17. The one exception to this statement is the case where we restrict the error correction term to be zero in the price and productivity equations. In this case, wage shocks have almost no long-run effect on prices. Phillips, A.W. 1958. "The Relationship between Unemployment and the Rate of Change of Money Wage Rates in the United Kingdom, 1861-1957." Econometrica, pp. 283-299. Phillips, Peter C. B. 1987. "Time Series Regressions with a Unit Root." Econometrica, pp. 277-301. Comovements among National Stock Markets Kenneth Kasa Economist, Federal Reserve Bank of San Francisco. I would like to thank Tim Cogley, Mark Levonian, and Carl Walsh for helpful suggestions. Barbara Rizzi provided excellent research assistance. International capital markets play an important role in the world economy. It is through these markets that risk and investment resources are allocated across countries. Gauging the extent to which international bond and equity markets perform these functions efficiently has therefore been a topic of great interest to economists. Traditionally, this question has been posed as whether or not national capital markets are "integrated" or "segmented." That is, do assets issued in different countries yield the same riskadjusted returns, or do they consistently yield different returns because of informational and governmentally imposed barriers? Clearly, if international capital markets are to provide appropriate signals to savers and investors, national bond and equity markets must be integrated. Attempts to answer this question are plagued by two difficulties not encountered in studies of domestic capital This paper uses the methodology of Hansen and Jagannathan (1991) to derive a lower bound on the correlation between any pair of asset returns under the hypothesis of complete markets. The bound is a simple function of the two assets Sharpe ratios and the coefficient ofvariation of a unique stochastic discount factor. The paper uses this bound to conduct robust1 nonparametric tests of the hypothesis that international equity markets are integrated. Using monthly stock return datafrom the U.S., Japan, and Great Britain for the period 1980 through 1993, I find that conclusions about market integration depend sensitively on the assumed variation of the (unobserved) common world discount rate. Given the observed correlations in returns, markets are more likely to be integrated the more volatile is the discount rate. market efficiency. First, assets issued in different countries tend to be denominated in different currencies, and exchange rate volatility adds an additional element of uncertainty to international investments. As a result, when testing the integration hypothesis one must either include a model of the pricing of exchange rate risk, or consider returns that have been "covered" against exchange rate risk. Second, because of taste differences and transportation costs, consumption patterns differ across countries much more than they do across regions within a single country. Since investors want to hedge their real consumption risks, this means that the riskiness of a given asset depends on the owner's country of residence. These problems make it even more difficult than usual to define a riskadjusted return, and consequently, make the results in this literature difficult to interpret. Studies of international bond markets generally conclude that markets are becoming increasingly integrated. This is particularly true when exchange risk and consumption differences are not an issue, e.g., when testing Covered Interest Parity. 1 Tests of Uncovered Interest Parity, however, have led to more ambiguous results. Although the hypothesis is typically rejected, no one has yet formulated 1. See Frankel (1993) for a survey of the evidence on short-term covered interest parity. Popper (1993) provides evidence on long-term covered interest parity. KAsA/COMOVEMENTS AMONG NATIONAL STOCK MARKETS an economic model of exchange rate risk that can explain these rejections. This has led some observers to question the efficiency of the foreign exchange market. 2 Even more stringent tests of international bond market integration, which require assumptions about both foreign exchange risk and international consumption differences, are conducted by Cumby and Mishkin (1986). They document close, but imperfect, linkages among the (ex ante) real interest rates of the U. S. and Europe. Glick and Hutchison (1990) apply the same methodology to real interest rate linkages between the U. S. and a set of Pacific Basin countries and find that financial liberalization has increased the linkages among these markets. In this paper, I examine the integration of international stock markets. Early work on this topic followed the same basic logic as bond market studies. That is, the extent of integration was judged by the correlation of returns, the idea being that greater equity market integration should lead to greater correlation among national stock markets. 3 Although this idea seems plausible, and in fact remains the conventional wisdom within the business community, we know. from the work of Lucas (1982) that the important implication of integrated capital markets is the equalization among countries of marginal rates of substitution in consumption, both intertemporally and across states of nature. Stock returns in an integrated market mayor may not be highly correlated, depending upon the nature of international specialization and the correlation of national productivity shocks. For example, stock markets may be segmented, yet stock returns could nonetheless be highly correlated if countries produce similar goods or if productivity shocks are highly correlated across countries. Conversely, stock markets might be integrated even if national stock returns are weakly correlated if countries are specialized in the production of different goods and if productivity shocks are weakly correlated across countries. This suggests that the coherence among national consumption growth rates probably provides a better metric for the degree of international capital market integration than does the correlation of stock returns. Obstfeld (1993) pursues this strategy and concludes that the weak relationships observed among national consumption growth rates are inconsistent with the hypothesis of internationally integrated capital markets, although he does find that markets have become more integrated over time. However, as Obstfeld himself acknowledges, this 2. Froot and Thaler (1990) survey the evidence on Uncovered Interest Parity. 3. See lorion (1989) for a survey of early work on international stock market integration. 15 approach suffers from a couple of severe drawbacks. First, in order to link consumption data to the marginal rate of substitution, one must specify a utility function. That is, this strategy is "parametric," and as a result one can never be sure whether a given rejection represents a bona fide rejection of the hypothesis of integrated markets or merely represents a rejection of the posited utility function. Second, it is widely recognized that consumption data contain measurement error. This creates econometric difficulties in implementing this approach. This paper adopts a strategy that avoids these problems. Not only is it nonparametric, and therefore robust to functional form misspecification and measurement error biases, but it also resurrects the intuitive notion that integration of equity markets should place restrictioJ:ls on the observed correlation among national stock markets. In particular, I adapt the methodology of Hansen and Jagannathan (1991) to derive a lower bound on the correlation between national stock market returns under the hypothesis of integrated markets. If the observed correlation between a pair of stock market returns is below its lower bound, then we can conclude that these markets do not share the same discount rate, or in other words, are not integrated. The basic idea behind this approach is as follows. Hansen and Jagannathan derive a lower bound on the volatility of an unobserved stochastic discount factor. This discount factor translates future state-contingent payoffs into current asset prices. Economic theories of asset pricing are distinguished according to how they link this discount factor to observable variables. For example, in the approach taken by Obstfeld the discount rate is assumed to be equal to the intertemporal marginal rate of substitution in consumption, while in the static CAPM it is assumed to be proportional to the return on the "market portfolio." Now, the hypothesis of integrated markets means that this discount factor is the same across countries, which implies that the Hansen-Jagannathan bound must be the same across countries. In particular, the lower bound on the standard deviation of the common world discount rate becomes a function of the observed variances and covariances of national stock market returns. In essence, all I do in this paper is invert this volatility bound to derive a lower bound on the correlation coefficient of returns as a function ofthe standard deviation of the unobserved discount factor. If the observed correlation is below this bound, then we must reject the joint hypothesis of integrated markets and the given value for the volatility of the stochastic discount factor. Before proceeding, one should understand the caveats to this approach. First, as always we are testing a joint hypothesis. This manifests itself here as the need to specify 16 FRBSF ECONOMIC REVIEW 1995, NUMBER 1 the standard deviation of the unobserved discount rate process. As we will see, we can always accept the hypothesis of integrated markets if we are willing to entertain a sufficiently volatile discount rate. The advantage of this approach,·therefore, is the flexibility it provides in linking the integration hypothesis to a broad spectrum of asset pricing models. We merely have to determine whether the volatility of the model-implied discount rate falls in a region that is consistent with the observed correlation of stock returns. If not, then either the discount rate model is false, or markets are segmented. The second caveat to keep in mind is that we are actually testing a stronger hypothesis than stock market integration. In particular, we are testing whether markets are complete, i. e., whether individuals have access to a full menu of dateand state-contingent securities, so that everyone, regardless of country of residence, has the same marginal rate of substitution in consumption, across all points in time and across all states of nature. Clearly, this is a very strong assumption. Stock and bond markets might be perfectly integrated, yet individuals could nonetheless end up with different marginal rates of substitution if these markets do not provide adequate insurance for all the risks that individuals face. Thus, as Obstfeld (1994) stresses in his recent survey, it would be desirable to develop a framework that allows us to test the stock market integration hypothesis without at the same time making such strong assumptions about the integration of goods markets and the nature of uncertainty. 4 The remainder of the paper is organized as follows. Section I briefly outlines the derivation of the HansenJagannathan bound on the volatility of stochastic discount factors. Section II then inverts this bound to get a lower bound on the correlation between asset returns. The correlation bound turns out to be a simple function of the two assets' Sharpe ratios and the coefficient of variation of the unobserved discount factor. Section III turns to empirical evidence. In particular, I consider whether the pairwise correlations among the stock markets of the U.S., Japan, and Great Britain satisfy their lower bounds. For standard models of the discount factor, observed correlations lie well below their lower bounds. This is because these models imply lower bounds that exceed unity. Of course, as noted above, rather than concluding that stock markets are segmented, an equally valid interpretation of this result is to reject the posited models of the discount factor. In fact, 4. In a related context, Tesar (1993, 1994) has stressed the need to incorporate nontraded goods into models of international capital market equilibrium. this has been the typical finding in this literature. 5 Not surprisingly, if we instead consider discount factors with volatilities approaching the Hansen-Jagannathan bounds reported in Bekaert and Hodrick (1992), we find that observed correlations satisfy their lower bounds. Finally, Section IV contains the conclusion and offers some suggestions for future research. I. DERIVING BOUNDS ON STOCHASTIC DISCOUNT FACTORS This section outlines how Hansen and Jagannathan (1991) use a set of observed asset returns to derive a lower bound on the volatility of an unobserved stochastic discount factor. The discussion will be brief, and the interested reader is urged to consult Hansen and Jagannathan's paper for full details. The starting point for the analysis is the following equation, which relates the price, 7T(P), of a given future state-contingent payoff, p, to an unobserved stochastic discount factor, m: 7T(P) = E(mp). (1) There are several ways to interpret this expression. The most general is to view m as the continuous linear pricing functional that is guaranteed to exist (by the Riesz Representation Theorem) as long as asset prices satisfy the "Law of One Price." If we also assume there are no arbitrage opportunities, then m must be nonnegative at all dates and in all states. Moreover, of particular relevance for this paper is the fact that if markets are complete, then m is unique (i.e., the same for all assets and all investors). While viewing m as an implication of the Riesz Representation Theorem provides a powerful unifying principle for asset pricing theories, a more intuitive interpretation of eq. (1) is to use the definition of the covariance operator to write it as follows: (2) 7T(P) = E(m)E(p) + cov (m,p) , Equation (2) illustrates the sense in which m plays the role of a discount rate. The first term on the right hand side of eq. (2) uses E(m) to discount the mean payoff, while the second term adjusts for the payoff's riskiness. Next, it often proves convenient to normalize asset prices to unity and rewrite eq. (1) in terms of asset returns: (3) 1 = E(mr) , 5. Employing standard utility function specifications, Obstfeld (1993) soundly rejects the consumption-based model of the discount factor. Frankel (1994) discusses the poor performance of static CAPM models of the discount factor. KASA / COMOVEMENTS AMONG NATIONAL STOCK MARKETS where r denotes the (gross) rate of return on an asset. Clearly, eq. (3) by itself imposes no restrictions on the data, since for a single asset we could always take m = 1/r. However, because the same m must satisfy eq. (3) for all returns, we have a set of overidentifying restrictions that can be tested if an explicit model for m is specified. This is the strategy pursued by Obstfeld (1993). However, to impose as little structure on the data as possible, Hansen and Jagannathan (1991) proceed nonparaInetrically and infer.bounds on the moments of m from the observed moments of a set of portfolio returns. To do this, note that since eq. (3) must hold for all assets (and, indeed, for all portfolios of assets), we can use the linearity of the expectations operator to subtract the analogous expression for the risk-free rate and get: (4) o= E(mre), where re denotes an asset's excess rate of return. Finally, define the n x 1 column vector of excess returns, Re, and write the vector analogue of eq. (4): (5) <T~ ~ WI13, where <T~ denotes the variance of m. Finally, from the algebra of least squares we know that (7) 13 = I-l [E(mRe) - E(m)E(Re)]. Using eq. (5), this can be simplified to: (8) Finally, plugging eq. (8) into eq. (6), and then rearranging, we get: (9) variation of the unobserved stochastic discount factor must be at least as large as the quadratic form on the right-hand side of (9). In the next section, I write out this quadratic form for the case of two stock returns, and then rearrange it to get a bound on their correlation coefficient as a function of <Tm/E(m). II. INVERTING THE HANSEN-JAGANNATHAN BOUND TO GET A CORRELATION BOUND The following proposition is the major result of this section. It relates the lower bound on the correlation between two assets to the two assets' Sharpe ratios and the volatility of a stochastic discount factor. PROPOSITION: Ifmarkets are complete, then the correlation between any pair ofexcess returns must satisfy the following lower bound: (10) p ~ x m2 -4(S.-S.)2 ' J l where Sj and Sj are the observed Sharpe ratios of assets i where m is a scalar, and 0 is an n X 1 column vector of zeros. Equation (5) provides a succinct representation of capital market equilibrium. Now, although m is not directly observable, imagine regressing m onto a constant and the vector of excess Re, where 0. is the regression returns, i.e., m = 0. + W intercept and 13 is the vector of slope coefficients. Of course, this regression will not provide a perfect fit. That is, there will be a regression error term, which by construction is uncorrelated with the fitted value from the regression. As a result, the variance of m must be at least as large as the variance of its predicted value. This variance is just equal to WI13, where I is the variance-covariance matrix of excess returns. In other words, it must be the case that (6) 17 ( <Tm E(m) . .. . an d j,. anu-l X m · is th"e coe).(fi' £Clent OJ+ variation OJ+ a unique (unobserved) stochastic discount factor. 6 The proof consists of two steps. First, with complete markets eq. (9) must hold (with the same unique m) for all collections of assets. By writing out the q~adratic form on the right-hand side of (9) for the case of just two assets, and then simplifying, we get: PROOF: (11) ( <T _m_ E(m) )2 == x2 m ~ 2s.s. __ ' J_ l+p + (1 +p)(l-p) Since this is nonlinear in p, it is convenient to take an approximation in order to be able to isolate p. Thus, the second step involves taking a first-order Taylor series expansion of 1/(1- p) around the point p = .5. Given the strict convexity of 1/(1 - p), this delivers the inequality 4p<1/(1- p). Using this in (11) and rearranging gives the bound in the proposition. Three points need to be made about this correlation bound. First, note that we could apply the quadratic formula in (11) and get a more precise bound. The only reason I take a linear approximation is to obtain a simple and easy to use expression for the bound. Calculations have shown that unless the true correlation bound is well outside the interval (0, 1) the approximation works quite well. )2 ~ E(Re)'I-lE(Re). Equation (9) is a version of Hansen and Jagannathan's volatility bound. It says that the (squared) coefficient of 6. The Sharpe ratio of an asset is its mean excess return divided by its standard deviation. 18 FRBSF ECONOMIC REVIEW 1995, NUMBER 1 Second, as noted earlier the bound contains no information that is not already contained in the volatility bound of Hansen and Jagannathan. In fact, plugging in the volatility bound implied by the two assets under consideration simply yields the actual observed correlation coefficient as the correlation bound! (Up to a second-order approximation error that is involved in deriving the bound.) Thus, the correlation bound is just an alternative expression for the Hansen-Jagannathan bound, although one that is perhaps more convenient to apply and interpret when assessing the market integration hypothesis. The third point to note is that the bound declines as the volatility of the discount factor increases. In other words, the more volatile is the discount rate implied by a given economic model, the more likely it is that the model will be consistent with the market integration hypothesis. At first it might seem puzzling that greater volatility in the discount rate lowers the required correlation between two stock markets. After all, the discount rate is a common factor in the stochastic evolution of the two markets, and increasing the variance of a common factor should increase the degree of comovement between the two markets. This intuition is indeed correct, but it ignores the distinction between covariance and correlation. Increasing the volatility of the discount rate also increases the standard deviation of stock returns, and it turns out that this offsets the greater covariance of returns, so that in the end the correlation bound decreases with CJ'm' m. EMPIRICAL EVIDENCE The previous section showed that the hypothesis of international stock market integration imposes restrictions on the correlations among national stock markets. Specifically, if markets are integrated, then the observed correlation between each pair of returns must exceed its lower bound given by eq. (10). If it doesn't, then we must either reject the posited value for the volatility of the discount rate, or conclude that these two markets do not share the same discount rate, and therefore are not integrated. Clearly, the crucial input in the analysis is the presumed volatility of the unobserved discount rate. From inspection of eq. (10), we can always accept the integration hypothesis if we posit a large enough value for CJ',jE(m). Thus, this section considers various specifications for this parameter and their associated implications for the hypothesis of international stock market integration. Before doing this, however, we must take a look at some data, since we also need to have values for the stock market correlations and the Sharpe ratios in each market. In this paper I consider the global economy's three major stock markets: those ofthe U.S., Japan, and Great Britain. The data are monthly, end-of-period observations on Morgan Stanley's CapitalInternational indices for the period 1980:1 through 1993:11. These indices are value-weighted and are based on a large sample of firms in each market. Crosslisted securities have been netted out. Returns are inclusive of (gross) dividend reinvestment and are expressed in U.S. dollar terms. Each return series is converted to "excess returns" by subtracting the one-month U.S. CD rate. Tne three pairwise correlations between excess returns are: (US, JP) = .246 (US, UK) = .531 (JP, UK) = .430 While the mean excess returns are (in units of percent per month): US = .563 JP = .847 UK= .747 And the standard deviations are: US = 4.46 JP = 7.42 UK = 6.30 Therefore, the Sharpe ratios tum out to be: US = .126 JP = .114 UK = .119 Thus, from the perspective of the static CAPM, it appears that over this period the U. S. market offered investors the best (dollar-denominated) risk/return trade-off. 7 Using these data, Figure 1 plots out the correlation bound for each country pair as a function of CJ',jE(m). The dotted line in each figure represents the actual observed value for the correlation coefficient. Evidently, of the three bivariate relations, the UK-Japan pair exhibits the strongest evidence in favor of market integration. That is, the bound is satisfied with the smallest value for CJ'm/E(m). On the other hand, the US-Japan pair exhibits the strongest evidence against the integration hypothesis, since it takes the largest value of CJ'm/E(m) to satisfy its correlation bound. Still, the differences are not large. Any value of CJ'm/E(m) below .135 indicates market segmentation, while a value greater than .155 would suggest market integration. What does economic theory tell us about the value of CJ'm/E(m)? As emphasized by Hansen and Jagannathan, traditional economic models ofthe discount rate have a hard time producing discount rates with this much volatility. For example, the two most commonly employed discount rate models are the static CAPM, which links the discount rate to the return on a (value-weighted) world market port- 7. As noted in the introduction, this inference can be misleading since residents of different countries consume different goods. As a result, it would probably be more accurate to consider own-currency excess returns, under the supposition that investors completely hedge the exchange rate risk associated with foreign equity investments. KASA / COMOVEMENTS AMONG NATIONAL folio, and the consumption-based CAPM, which links the discount rate to a representative agent's intertemporal marginal rate of substitution in consumption. Constructing a value-weighted portfolio from the U.S., Japanese, and British markets produces a value for CTm/E (m) of just.045. Moreover, unless you introduce habit persistence or statenonseparabilites, the consumption-CAPM produces even smaller values for CT,jE(m). Values for CT,jE(m) of this order of magnitude then produce lo~ver correlation bounds well in excess of unity. By itself, this suggests that international stock markets are segmented. However, remember that an alternative interpretation of these results is to reject the posited models of the discount rate. How are we to decide between these two inferences? While the joint nature of the integration hypothesis can never be eluded entirely, the approach in this paper facilitates the choice between the two interpretations. Clearly, it makes no sense to test the hypothesis of international stock market integration using a model for the (common) discount rate that produces a value for CTm/E(m) that is below the maximum Sharpe ratio of the considered countries. After all, two markets cannot be "integrated" if each individually violates the Hansen-Jagannathan bound. 8 In the context of this paper, this means that it only makes sense to consider discount rate models with implied volatilities (i.e., coefficients of variation) in excess of .126, which is the maximal Sharpe ratio among the markets of the U.S., Japan, and Great Britain. Testing the integration of these three markets using a model that implies a less volatile discount rate is bound to lead to ambiguous results, as it is not even consistent with domestic stock market efficiency. Figure 1 illustrates that it is possible to conclude that individual national stock markets are efficient, but not integrated. For example, any value of CT,jE(m) that lies between .13 and .15 is a viable model of the U.S. and Japanese equity markets considered in isolation, but is inconsistent with the hypothesis that these two markets are integrated. As noted earlier, however, it is quite difficult to formulate economic models with implied discount rate volatility anywhere near .13. This suggests that parametric testing of the international stock market integration hypothesis must await further advances in dynamic asset pricing theory. Given this state of affairs, the nonparametric approach of this paper provides a valuable tool for the assessment of international stock market integration. Specifically, by SroCK MARKETS FIGURE 1 CORRELATION BOUNDS JAPAN AND 0'8 K 0.7 _ _. 0.6 U.S. Lower Bound 0.5 0.4 0.3 0.2 0.1 o +---,---,-----,--,---..,..---, 0.13 U. S. 0.135 0.14 0.145 0.15 0.155 0.16 AND GREAT BRITAIN 0.8~ 0.7 Accept ---1"~ Reject... 0.6 ~ ~ __ 0.5 0.4 0.3 0.2 0.1 0 0.13 0.135 0.14 0.145 0.15 0.155 0.16 0.155 0.16 JAPAN AND GREAT BRITAIN 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.13 8. By the Cauchy-Schwarz inequality, adding countries to the calculation of the Hansen-Jagannathan bound must increase(or at least not decrease) the lower bound on the volatility of the discount factor. 0.135 0.14 0.145 0.15 NOTE: The horizontal axis measures the coefficient of variation of the discount factor. 19 20 FRBSF EcONOMIC REVIEW 1995, NUMBER 1 checking to see by how much the minimum integrationconsistent discount rate volatility exceeds the maximal Sharpe ratio, we get a rough idea of how likely it is that two markets are integrated. This is the basis for the previous conclusion that the integration hypothesis is most strongly supported in the case of Japan and Great Britain, and least supported in the case of the U.S. and Japan. 1 V. CONCLUSIONS AND EXTENSIONS This paper has developed a simple, "back-of-the-envelope" procedure for determining whether the observed correlation between two national stock markets is consistent with the hypothesis of international stock market integration. All you have to do is get data on the countries' Sharpe ratios, and then specify a parameter that captures the volatility of the common world discount rate. The advantage of the approach is that it provides a flexible and intuitive method for mapping out the relationship between economic theory (as expressed in a discount rate model) and normative conclusions about capital market efficiency. Perhaps not surprisingly, conclusions about market integration depend sensitively on this parameter. The more volatile is the posited discount factor, the more likely it is that observed comovements among national stock markets are consistent with the hypothesis of internationally integrated markets. As noted above, none of the standard economic models of asset pricing produce discount factors that are sufficiently volatile to be consistent with the hypothesis of integrated markets. However, I argued that at this stage it is more appropriate to reject the models than it is to reject the integration hypothesis. In addition, I argued that the nonparametric approach of this paper provides guidance on how to construct models that are minimally capable of addressing the integration hypothesis, in the sense that they are at least viable models of domestic capital market efficiency. As a final caveat, one should keep in mind that all of the analysis here has been based on point estimates that are subject to sampling variability. Future work along these lines should attempt to incorporate standard statistical inference considerations. This would enable us to assess the statistical significance of observed differences between market correlations and their lower bounds. The recent work of Hansen, Heaton, and Luttmer (1993) should provide the necessary tools for such an extension. REFERENCES Bekaert, Geert, and Robert 1. Hodrick. 1992. "Characterizing Predictable Components in Excess Returns on Equity and Foreign Exchange Markets." Journal ofFinance 47, pp. 467-508. Cumby, Robert E., and Frederic S. Mishkin. 1986. "The International Linkage of Real Interest Rates: The European-US. Connection." Journal ofInternational Money and Finance 5, pp. 5-23. Frankel, Jeffrey A. 1993. "Quantifying International Capital Mobility in the 1980s. in On Exchange Rates (Chapter 2). MiT Press. _ _ _ _ . 1994. "Introduction." In The Internationalization of Equity Markets, ed. 1. Frankel. NBER. Froot, Kenneth A., and Richard H. Thaler. 1990. "Anomalies: Foreign Exchange." Journal of Economic Perspectives 4, pp. 179-192. Glick, Reuven, and Michael Hutchison. 1990. "Financial Liberalization in the Pacific Basin: Implications for Real Interest Rate Linkages." Journal of the Japanese and International Economies 4, pp. 36-48. Hansen, Lars P., John Heaton, and Erzo Luttmer. 1993. "Econometric Evaluation of Asset Pricing Models." NBER Technical Working Paper No. 145. Hansen, Lars P., and Ravi Jagannathan. 1991. "Implications of Security Market Data for Models of Dynamic Economies." Journal of Political Economy 99, pp. 225-262. Jorion, Philippe. 1989. "The Linkages between National Stock Markets." In The Handbook ofInternational Financial Management, ed. R. Aliber. Dow Jones-Irwin. Lucas, Robert E., Jr. 1982. "Interest Rates and Currency Prices in a Two-Country World." Journal of Monetary Economics 10, pp. 335-359. Obstfeld, Maurice. 1993. "Are Industrial-Country Consumption Risks Globally Diversified?" NBER Working Paper No. 4308. _ _ _ _ . 1994. "International Capital Mobility in the 1990s." In Understanding Interdependence: The Macroeconomics of the Open Economy, ed. Peter Kenen. Princeton Univ. Press. Popper, Helen. 1993. "Long-Term Covered Interest Parity: Evidence from Currency Swaps." Journal of International Money and Finance 12, pp. 439-448. Tesar, Linda L. 1993. "International Risk-Sharing and Nontraded Goods." Journal ofInternational Economics 35, pp. 69-89. _ _ _ _ . 1994. "Evaluating the Gains from International RiskSharing." Unpublished manuscript. Changes in the Structure of Urban Banking Markets in the West Elizabeth S. Laderman Economist, Federal Reserve Bank of San Francisco. The author wishes to thank Jennifer Catron and Deanna Brock for research assistance. This paper begins with a discussion of the influence of the number of firms and the variance of market shares on the Herfindahl-Hirschman Index (HH/) measure of market concentration. The paper then reports the changes in the number of depository institutions (DIs) and in the HHI in the Twelfth District and its 65 individual urban banking markets between 1982 and 1992, attributing these changes to underlying causes. I find that, although an increase in concentration need not accompany a decrease in firms, more than two-thirds of the 53 markets with DI decreases showed concentration increases. This suggests that regulatory review of DI mergers has been and will continue to be important in assuring the competitiveness of banking markets. Over the past decade, consolidation has led to many changes in the banking landscape. In the West, mergers such as those between Wells Fargo Bank and Crocker National Bank and between Bank of America and Security Pacific National Bank, as well as many less dramatic combinations, have eliminated banks from Alaska to Arizona. At the same time, numerous banks and even more thrifts have failed. Although brand new banks and thrifts continue to be formed, between 1982 and 1992, the Twelfth District saw the number of depository institution competitors decline by 15 percent, from 932 to 792,1 In this paper, I will discuss the changes wrought by a decade of bank and thrift mergers, failures, and entry on the structure of urban banking markets in the West. 2 Market structure is important because it is thought to influence competition, which, ultimately, can affect the welfare of the entire economy. The paper will focus on two aspects of market structure: the number of competitors and the concentration of market shares. The paper will proceed as follows. In the first section, I briefly discuss the structure-conduct-performance paradigm of industrial organization theory. I also introduce the concept of market concentration and the statistic often used to measure it, the Herfindahl-Hirschman Index (HHI). The second section discusses how changes in the distribution of market shares and the number of depository institutions affect concentration. In the third section, I discuss the changes in concentration and in the number of depository institutions in the Twelfth District overall. In the fourth section, I report the changes in concentration and in the number of depository institutions in 65 local urban markets between 1982 and 1992 and attribute these changes to underlying causes. I also draw some general conclusions 1. Here and throughout the paper, I refer to the states ofAlaska, Arizona, California, Hawaii, Idaho, Nevada, Oregon, Utah, and Washington collectively as the Twelfth District. The Federal Reserve Bank of San Francisco, which serves the Twelfth Federal Reserve District also serves American Samoa, Guam, and the Commonwealth of the Northern Mariana Islands. 2. Relatively little research of this type has been conducted. However, David Holdsworth (1993) does provide some information on changes in the structure of banking markets in New York and New Jersey between 1980 and 1991. 22 FRBSF ECONOMIC REvIEW 1995, NUMBER 1 regarding causes for changes in concentration and competitiveness in these markets. The fifth section concludes the paper. 1. THE STRUCTURE-CONDUCT-PERFORMANCE PARADIGM AND THE HHI The structure-conduct-performance paIadigm states that market structure influences firm conduct and, in tum, economic performance, and that the direction of such effects often is predictable. Elements of market structure include the number and size distribution of sellers and buyers, the degree of product differentiation, and the existence and extent of barriers to entry into the market. Characteristics of firm conduct include pricing behavior, advertising strategy, and technological innovation. Performance includes the efficiency of production and resource allocation. 3 It is one of the most fundamental structure-conductperformance theories of industrial organization that the smaller the number of firms dominating a market, the more likely those firms will be able to collude to maintain prices above the competitive level and thereby operate at an inefficient point on the production function. This link makes the study of market concentration valuable. Ideally, one would study market performance directly, because this is what we really care about, but this usually is infeasible. For example, determining the efficiency of production requires knowing the production technology, which often is very difficult, especially for multidimensional services such as banking. Alternatively, the structure-conductperformance paradigm suggests that the conduct of firms is closely connected to performance. Here again, however, we often cannot directly observe firms' behavior. However, market concentration usually is fairly easy to measure. Determining changes in market concentration, then, can help to suggest the changes that may have taken place in competitiveness and productive efficiency. However, the relationship between concentration and efficiency is not necessarily as unambiguous as just described. Many economists have pointed out that more concentrated markets may in fact be more efficient. This could be because efficient firms are more profitable, which causes them to grow and acquire market share. Therefore, efficient markets are ones in which there are a few large, profitable, and efficient firms, and inefficient markets are 3. EM. Scherer discusses the structure-conduct-perfonnance paradigm and uses it as the organizing theme for his classic textbook of industrial organization theory, Industrial Market Structure and Economic Performance (1980). ones in which no such efficient firms have emerged to take the lead in market share. In addition, economies of scale or scope in production may mean that being large causes a firm to be efficient. Again, this may mean that a market with a few large firms is more efficient than a market with many smaller firms. Despite the validity of these arguments, numerous empirical studies support the view that, in many industries and under many circumstances, the greater the concentration of output in a small number of firms, the greater the likelihood of welfare losses due to weak competition and thereby low efficiency of production. The Herfindahl-Hirschman Index In fact, the link between concentration and the likelihood of welfare losses is sufficiently accepted that an assessment of the change in concentration is central to regulators' analyses of the effects of proposed mergers between firms in many industries, including banking. Banks and bank holding companies must apply to one or more ofthe federal banking agencies for approval of mergers and acquisitions. If regulators find that the proposed transaction would raise market concentration too much, the merger or acquisition application may be denied, or divestitures of branches or other assets to third parties may be required. To measure concentration in banking markets, the federal bank regulatory agencies and the Department of Justice (DOJ) use a statistic called the Herflndahl-Hirschman Index (HHI). The HHI is computed as the sum of the squares of the percentage of deposits held by each of the competitors in a market. For example, if a market has only one firm, then the HHI is 1002 , or 10,000. If the market is evenly divided between two firms, the HHI is 502 + 502 , or 5,000. The following example illustrates the use of the HHI for the evaluation of a hypothetical merger. Say that there are four banks in a market: A, B, C, and D. Say that, before the merger, A produces 35 percent of the output in the market, B 30 percent, C 20 percent, and D 15 percent. The premergerHHI, then, is35 2 + 302 +202 + 15 2 ,or2,750. Now, assume that banks B and D merge. The HHI after the merger would be 35 2 + (30+ 15)2+ 202 , or 3,475. The merger inreases the HHI by 725. For evaluating individual mergers, the DOl's bank merger policy indicates that a bank merger that increases the HHI in a local market by 200 points or more and results in an HHI of at least 1,800 would raise competitive concerns. While the policy is not hard and fast, its use has led to the denial of merger applications and, more often, to the divestiture of banking offices to third parties to reduce the effects on market concentration. As a result, the policy has LADERMAN / STRUCTURE OF URBAN BANKING MARKETS IN THE WEST helped contain the adverse effects of individual mergers on competition. In addition to being used for individual merger analysis, the HHI can be used to track changes in concentration over a period of time. Changes in concentration may be due to mergers, acquisitions, failures, withdrawals from the market, or simple shifts in market shares due to the dynamics of competition among an established set of banks and thrifts. The purpose of this paper is to describe how concentration and competitiveness in urban banking markets in the West changed between 1982 and 1992 and to discuss underlying causes for these changes and implications for policy. Because of its use in the competitive analysis of bank mergers, the HHI is an intuitively appealing measure of concentration and will be used in this paper. 4 ll. DETERMINANTS AND DYNAMICS OF THE-HIll In this section, I will discuss the relationship between the HHI and its two underlying determinants: the number of firms in a market and the distribution of market shares among those firms. The key to this relationship is the recognition that the HHI can be decomposed into the sum of two terms, one that depends on the number of firms and one that depends on the variance of their market shares. 5 The HHI is given by N HHI = ;~1 (1) xl, where Xi is the percentage market share held by firm i and N is the total number of firms in the market. The variance of market shares, V, is 1 V = N (2) N ! (x. ;=1 - x)Z , 1 where x is the mean market share. Noting that I V = - (3) N x= 1 - XZ 1 ' N N ! x ;=1 i -- 100 N ' we have, from (1), (5) HHI Equation (5) states that the HHI is the sum of two terms, the first a function of the number of firms and the variance of market shares and the second a function only of the number of firms. Two conclusions emerge directly from equation (5). First, the HHI increases with the variance of market shares. Therefore, given the number of firms, if the variance is at its minimum, the HHI also must be at its minimum. The minimum value of the variance is zero, and this yields a minimum HHI of lO,OOO/N. Second, if the HHI exceeds 1O,000/N, it must be because the variance is greater than zero. By definition, the variance of a group of numbers is zero if and only if all of the numbers are equal. Therefore, the first term on the right-hand side of equation (5) can be interpreted as the contribution to the HHI ofthe dispersion of market shares away from equality, the "inequality effect," while the second term is what the HHI would be were the market shares of all N firms equal, the "number of firms" effect. Because the HHI depends on the variance of market shares, shifts in the distribution of market shares affect the HHI. The effect of a change in market shares, holding N constant, can most easily be seen for the case in which only two market shares change. Let the original HHI be given by (1). Then, let the new HHI be given by N (6) HHI' = yZ + zZ + ! x.Z • ;=3 1 Here, in the new distribution of market shares, Xl and Xz have been replaced by y and z, but none of the other market shares have changed. Subtracting (1) from (6), one finds that the HHI rises if and only if (7) yZ + ZZ > x1Z + xl . However, we know that the sum of market shares must always be 100, so we can use the requirement that Xl + Xz equal y + z, and therefore that their squares be equal, to simplify the above condition to (8) N ! x.z ;=1 and that (4) 23 = NV + Nx Z = NV + 10,000 N 4. For more on the HHI, see Rhoades (1993). 5. I thank Mark Levonian for pointing out this relationship. We now see that, if two of the market shares change, the new HHI will exceed the old HHI if and only if the product of the new shares is less than the product of the old shares. The product of two numbers, the sum of which is a constant, increases as the two numbers converge and decreases as they diverge. Therefore, the new HHI will exceed the old HHI if the two shares have diverged and will be less than the old HHI if the two shares have converged. This also is the condition under which the variance of market shares increases when two market shares change. This is expected: From (5), it is apparent that, if the number of firms is held constant, concentration increases if and only if the variance of market shares increases. 24 FRBSF ECONOMIC REvIEW 1995, NUMBER 1 The insight offered by equation (8) also provides a convenient way to prove that within-market mergers must increase concentration if none of the market shares of the uninvolved firms change. Let N be the number of firms in the market before the merger, and let the firms that will merge, firm I and firm 2, have market shares of Xl and x2 . After the merger, one can still think of the market as having N firms. The new, merged firm, has market share y=xI + x 2 ' and it can be thought to have repiaced, say, firm L Firm 2's new share, z, is now zero. As long as the market shares of all of the N - 2 uninvolved firms have not shifted, concentration must have increased because the shares of the involved firms have diverged. 6 Likewise, the entry of a new firm into a market must decrease concentration as long as the market share of only one firm already in the market is affected. The condition under which the HHI will increase when more than two market shares change is a simple generalization of the condition expressed in (8): In order for any number of changes in market shares to increase the HHI, the sum of all of the cross-products of the new market shares has to be less than the sum of all of the crossproducts of the shares that they replaced. (This is exactly the condition under which the variance of market shares increases when market shares change.) For example, if, in two lists .of equal numbers of market shares, three market shares differ across lists, the HHI for the new list will be greater than the HHI for the old list if and only if (9) X IX2 + X I X3 + Xr3 > wy + wz + yz, where w, y, and z are the new market shares, and Xl' x2 , and x 3 are the shares that they replaced. A final point regarding the relationship between the HHI and the distribution of market shares is that, using (5), (to) aHHI av = N. Holding N constant, a given increase in the variance increases the HHI more, the greater the number of firms. Regarding the relationship between the number of competitors and the HHI, equation (5) offers several insights. It says that the minimum HHI, obtained when the variance is zero and all market shares are equal, is lower with more firms in the market (higher N). In addition, equation (5) provides intuition for the meaning of the DOl's definition 6. There is a tendency for some acquiring banks to lose some of the combined market share of the merged firms following an acquisition. Sometimes, competitors have been able to attract customers from merged institutions because they closed branches or otherwise changed bank practices. This type of effect helps to reduce the concentrating effects of within-market mergers. of a "highly concentrated" banking market. 7 Express any value of the HID as 10,000 times the inverse of some number. Then, that number is the number of equal-sized firms that would give the same value of the HHI. The DOJ definition of a highly concentrated banking market as one with an HHI of at least 1,800 means that a market with six equal-sized banks is not too concentrated, but one with five equal-sized banks is. This is because 10,000 10,000 (11) - 6 - = 1,666.67 < 1,800< - 5 - = 2,000. If market shares are not equal, the relationship between the number of firms and the HHI is somewhat more complicated. If V is held constant, we can determine the effect on the HHI of an increase in the number of firms by taking the partial derivative of the HHI with respect to N. From (5), we have aHHI (12) aN 10,000 = V - ~ Holding V constant, an increase in N lowers the HHI if V is less than 10,0001N2 and raises the HHI if V is greater than 1O,OOOIN2. Also, note that (13) a2H aN2 = 2 N3 > 0, so that, the larger is N, the less the decline in the HHI when firms are added. In addition, note that the first term on the right-hand side of (12), the partial derivative of the inequality effect with respect to N, is positive as long as V is positive. This means that, as long as V is positive, an increase (decrease) in the number of firms will increase (decrease) the effect of the inequality of shares on concentration, even if inequality as measured by the variance does not change. However, it is likely that, in many situations, equations (12) and (13) do not apply. This is because, in practice, a change in the number of firms must change some market shares and therefore likely will change the variance of market shares. If the variance changes, the derivative of the HHI with respect to N is given by: (14) dHHI dV dN = V + N dN - 10,000 ~ Unfortunately, the variance of market shares can change any number of ways as the number of firms changes, so neither the sign nor the size of dVl dN is known. 7. The DOJ classifies markets with an HID below 1,000 as "unconcentrated," those with an HHI between 1,000 and 1,800 as "moderately concentrated," and those with an HHI above 1,800 as "highly concentrated." 25 LADERMAN / STRUCTURE OF URBAN BANKING MARKETS IN THE WEST However, it is straightforward to derive an expression for the discrete change in the HHI in terms of given discrete changes in N and V. Using equation (5), the change in the HHI due to moving from initial levels No and Vo to levels N 1 and VI is: (15) aBBI = BBI1 = N1V1 + BBIo 10,000 -NoVo - aBBI = N1a V + (VO - EFFECTS ON HHI OF CHANGES IN INEQUALITY EFFECT, VARIANCE, AND NUMBER OF FIRMS PANEL N Increases ~. o 10,000 NoN )aN. NV Increases NV Decreases Here, one can see that, as long as the initial variance of market shares is greater than 1O,000INoNI' an increase in the number of firms along with an increase in the variance of market shares guarantees that concentration will increase. On the other hand, if initial variance exceeds 1O,000/NoNl and V decreases, an increase in the number of firms will not necessarily increase concentration. The condition that initial variance exceed 1O,OOOINoNl is the discrete analogue to the condition in equation (12) that initial variance exceed 10,000/!V2 in order for an increase in N to increase concentration if the variance of market shares does not change. If initial variance is less than 10,0001NoNI' then an increase in N along with a decrease in V definitely will lower concentration. However, if V increases under these circumstances, concentration may increase. Equation (5) also yields an alternative decomposition of discrete changes in concentration. Simply, aBHI = (N1V1 - NoVo) 10,000 10,000 + ( N - ~). I 0 The first term in (17) is the change in the inequality effect, and the second term is the change in the numbers of firms effect, that is the change in concentration in going from No equal-sized firms to N 1 equal-sized firms. Equivalently, it is the change in concentration in going from No firms to N 1 firms, while holding the inequality effect constant. Clearly, the second term is positive if and only if N 1is less than No. Of course, the changes in the variance of market shares, the number of firms, and the inequality effect all interact with one another, and one can combine the two decompositions in (16) and (17). Table 1 shows what happens to concentration given various combinations of increases and decreases in the number of firms, the inequality effect, and the variance of market shares. Panel A gives the breakdown for the case in which the initial variance of No Change in N N Decreases V Increases + + + V Decreases + or- n.a. n.a. V Increases n.a. n.a. + or- V Decreases I (17) A: Vo > (lO,OOO/NoNl) 10,000 Subtracting and adding N1Vo and gathering terms, this yields: (16) TABLE 1 PANEL B: Vo < (lO,OOO/NoNl) N Increases NV Increases V Increases + or- V Decreases NV Decreases V Increases V Decreases n.a. No Change in N N Decreases + + n.a. n.a. n.a. + + or- NOTE: NV = inequality effect V = variance N = number of firms market shares is greater than 1O,0001NoNl and Panel B gives the breakdown for the case in which it is less than 1O,000INoNl. Some of the cases in the table have ambiguous implications for concentration. The decomposition given in this table will be used to show the underlying causes for increases and decreases in concentration in local banking markets in the Twelfth District. To summarize the important conclusions of this section: 1. Concentration depends on a "number of firms effect" and an "inequality effect," so changes in concentration depend on changes in these factors 2. The inequality effect itself depends on the number of firms and the variance of market shares 3. When both the number of firms and the variance of market shares change, the change in concentration depends on changes in these factors and on the size of the initial variance of market shares relative to a function of the initial and terminal numbers of firms. 26 FRBSF ECONOMIC REVIEW 1995, NUMBER 1 ill. CHANGES IN THE NUMBER OF DEPOSITORY INSTITUTIONS AND IN THE HHI IN THE TWELFTH DISTRICT The number of bank and thrift competitors in the Twelfth District declined by approximately 15 percent between 1982 and 1992, as the number of bank competitors went from 631 to 612, and the number of thrift competitors went from 301 to 180. 8 f'.~ote that these are the numbers of separate bank and thrift competitors, not the numbers of banks and thrifts. Many banks and some thrifts are subsidiaries of holding companies, and some of these holding companies have more than one bank or thrift subsidiary. Because they have common corporate control, I do not count separate subsidiaries of the same holding company as separate competitors. 9 I will refer to bank and thrift competitors as depository institutions (DIs). The number of DIs is the sum of all bank and thrift holding companies plus the number ofbanks and thrifts that are not holding company subsidiaries. The number of DIs in the District has been influenced by several forces. First, there have been mergers between DIs. A merger between an in-District DI (a DI with at least one branch in the District, but not necessarily headquartered in the District) and another in-District DI, or an acquisition of an in-District DI by an in-District DI, reduces the number of DIs in the District by one. When the assets and liabilities of one DI are split up and sold to multiple DIs, this also reduces the number of DIs by one. Some mergers or acquisitions may have involved an out-of-District DI merging with or acquiring an in-District DI. These would have only changed DIs' names and would not have affected the number of DIs in the District. Second , some DIs have failed, and their assets and liabilities have been taken over by other DIs. In essence, these were acquisitions, although it is likely that many of them differed from ordinary acquisitions in that the buyer received government assistance for the purchase. However, some DIs failed and were completely liquidated, with insured depositors paid off by the bank or thrift deposit insurance fund. Each such failure reduced the number of DIs by one. Finally, some new DIs came into being, and each occurrence raised the number of DIs by one. New DIs arise when an applicant receives a new bank or thrift charter. Note, however, that when a holding company already in the 8. Only bank and thrift organizations that held deposits in these years were counted. 9. This is consistent with the practice followed by the federal depository institution regulatory agencies in the analysis ofthe competitive effects of bank and thrift mergers and acquisitions. market establishes a new subsidiary bank or thrift in the market by obtaining a new charter, this does not change the number of DIs in the market. On the other hand, ifeither an out-of-market holding company or a completely new entity obtains a new charter and sets up a new bank or thrift in the market, this raises the number of DIs in the market by one. lO The actual numbers of mergers, failures, and new entries of DIs in the Threifth District between 1982 and 1992 are somewhat difficult to pinpoint. For example, it is much easier to determine the number of bank and thrift mergers than the number of DI mergers and acquisitions, and the two are not necessarily the same. Two banks that are subsidiaries of the same bank holding company may merge, but this does not change the number of DIs. In addition, a bank holding company may merge with another bank holding company, and each of several target banks may be merged into separate subsidiaries of the surviving bank holding company. Such a transaction would eliminate only one DI, even if it generated several bank mergers. On the other hand, a· bank holding company can acquire a bank without merging it into another bank, and a list ofbank mergers would not include such acquisitions. If the acquired bank was not part of a holding company, or if its former holding company had only one bank, this reduces the number of DIs by one. If the acquired bank was part of a holding company that still has at least one bank subsidiary after the acquisition, the acquisition does not affect the number of DIs. Other complications involve the number of DI failures and new entries. It is fairly straightforward to determine the number of liquidated banks and thrifts. However, some of the liquidated banks or thrifts may be subsidiaries of holding companies with other still solvent subsidiaries, in which case the disappearance of the bank or thrift does not constitute the disappearance of a DI. Finally, one can easily determine the number of new bank and thrift charters granted between 1982 and 1992, but it is much more difficult to know whether or not those charters were granted to existing DIs. 10. Branching by established out-of-market DIs also can increase the number of DIs. Also, acquisition of only some of the branches of a DI in a market by an out-of-market DI will increase the number of DIs by one. Most states in the District pennit nationally chartered out-of-state thrifts to branch into their state by setting up new branches or acquiring existing branches, but only two states in the District pennit interstate branching by banks. Utah pennits out-of-state banks to operate offices in Utah as branches, and Nevada pennits out-of-state banks to set up new branches in Nevada counties with a population less than 100,000. However, it is likely that any interstate thrift or bank branching would have had a very minor effect on the change in the number of DIs in the District as a whole. LADERMAN / STRUCTURE OF URBAN BANKING MARKETS IN THE WEST Given the above complications, the following numbers of bank and thrift liquidations, new charters, and mergers will only approximate the number of banking competitor and thrift competitor liquidations, mergers, and new formations. There were 13 bank liquidations and 13 thrift liquidations in the District between 1982 and 1992. There were 113 new thrift charters granted and 324 new bank charters. Finally, there were 333 mergers in which the acquirer was a Twelfth District bank and the target was a Twel~h District bank or thrift and 208 mergers in which the acquirer was a Twelfth District thrift and the target was a Twelfth District bank or thrift. Subtracting the total number of bank liquidations and mergers from the 631 bank competitors existing in 1982 and adding the number of new bank charters yields 609 bank competitors, which is close to but slightly less than the actual number in 1992, 612. Subtracting the total number of thrift liquidations and mergers from the 301 thrift competitors existing in 1982 and adding the number of new thrift charters yields 193 thrift competitors, which is close to but somewhat greater than the actual number in 1992, 180. However, these estimates of the changes in the numbers of bank and thrift competitors, obtained by using the above numbers for failures, mergers, and new entries, is close enough to the actual change that two conclusions seem warranted. First, the complete disappearance of DIs through failure likely was relatively uncommon between 1982 and 1992. Many failing banks and thrifts may have been eliminated by way of merger or acquisition, but few were entirely liquidated. Second, the decrease in the number of DIs between 1982 and 1992 was caused by a massive number of mergers and acquisitions (about 541) that was not quite balanced by the very large number of new entries (about 437). The 15 percent net decrease in the number of DIs between 1982 and 1992 may be considered to be relatively modest, but the large gross numbers suggest that the underlying forces causing that decrease likely were not. In addition, the disappearance of thrift competitors accounted for a much larger proportion of the net decrease in DIs than did the disappearance of bank competitors. Over the ten-year period, on net 121 thrift competitors disappeared, accounting for 86.4 percent of the 140 DIs eliminated on net. As discussed in the last section, within-market mergers and acquisitions must raise market concentration if the notinvolved firms' market shares do not change. On the other hand, unless shifts in the market shares of more than one preexisting bank accompany new entry, new entry will lower concentration, thereby increasing the likelihood of vigorous competition. It appears that there were more DI mergers and acquisitions than new entry of DIs between 1982 and 1992 in the Twelfth District. Therefore, taking 27 into account only the changes in the number of DIs and not any shifts in market shares among existing competitors, it is likely that banking market concentration in the Twelfth District as a whole increased between 1982 and 1992. To investigate this possibility, I calculated HHls for 1982 and 1992 for the banking and thrift industry for the entire Twelfth District. I calculated the HHI the same way that the Federal Reserve does in its analysis of the competitive effects of D1 mergers. Specifically, each DI's market share is the percent of total market deposits (in this case, total deposits in the Twelfth District) that it holds. In addition, thrifts are considered to be only partial competitors of banks. This is because thrifts usually are prohibited from engaging in all of the activities in which banks participate. For example, thrifts' commercial lending often is restricted. Therefore, it is customary to give only a 50 percent weight to thrift deposits when calculating the size of the market and market shares. 11 For example, say that a market is comprised of two banks and a thrift. The first bank has $500 million in deposits, the second bank has $300 million, and the thrift has $300 million. Weighting the thrift deposits at 50 percent and the bank deposits at 100 percent, total deposits in the market are $950 million. The first bank's percent market share is 52.6 percent, the second bank's share is 31.6 percent, and the thrift's share is 15.8 percent. Summing the squares of these market shares yields an HHI of 4,015. Using deposits to measure market share and applying a 50 percent weight to thrift deposits, the HHI for the Twelfth District did indeed rise between 1982 and 1992, from 586 to 820. Apparently, the inequality effect either increased or did not decrease enough to outweigh the effect of the net decrease in DIs on concentration in the Twelfth District. This suggests that the competitiveness and productive efficiency of banking in the Twelfth District fell between 1982 and 1992. 11. When a bank merges with or acquires a thrift, the pre-merger calculation of the HHI weights all thrift deposits at 50 percent, but the post-merger calculation of the HHI weights the merged or acquired thrift's deposits at 100 percent and the other thrifts' deposits at 50 percent. This procedure reflects the post-merger control ofthe acquired thrift's deposits by a banle When a bank merges with or acquires another bank, both the pre- and post-merger calculations of the HHI weight all thrift deposits at 50 percent. Consistent with this, all HHIs and total deposit figures that are reported in this paper were derived by applying a 100 percent weight to all bank-controlled deposits and a 50 percent weight to all thrift-controlled deposits. Specifically, if a bank holding company has a thrift subsidiary, that thrift's deposits are weighted at 100 percent, not 50 percent. This, however, is relatively unusual. 28 FRBSF ECONOMIC REVIEW 1995, NUMBER 1 IV CHANGES IN THE NUMBER OF DEPOSITORY INSTITUTIONS AND IN CONCENTRATION IN LoCAL URBAN MARKETS My ultimate focus is on changes in the level of competition between banking organizations, and therefore changes in concentration in meaningfully defined banking markets are more important than changes at the District level. Because many banking services are supplied locally, and many bank customers find it very costly to look for alternatives outside their local area, the antitrust analysis of bank mergers typically defines banking markets to be local. Therefore, I investigated changes in the structure of 65 local urban banking markets between 1982 and 1992 in the Twelfth Federal Reserve District states of Alaska, Arizona, California, Hawaii, Idaho, Nevada, Oregon, Utah, and Washington. 12 These urban banking markets are geographically defined to correspond to Rand McNally's "RaNally Metro Areas," or RMAs. The geographic boundaries of RMAs are delineated by Rand McNally to include the areas around important cities that are developed and economically integrated with the urban center. RMAsjnclude satellite communities and suburbs as well as one or more centralcities.l3 Every RMA in the Twelfth District is represented in my urban banking market sample. Most of the Twelfth District population lives in RMAs, and most of the DI deposits reside in branches in RMAs. In 1980,86.8 percent of the Twelfth District population lived in RMAs, and, in 1990, 86 percent lived in RMAs. In 1982, approximately 88.9 percent of the total deposits in the Twelfth District were held in branches located in RMAs, and in 1992 this percentage was about 88.3. Tables 2a and 2b present rank order listings of the RMAs by the change in HHI between 1982 and 1992; 2a is ordered by HHI increases and 2b is ordered by HHI decreases. As described above, the HHls are calculated using 100 percent of bank deposits and 50 percent of thrift deposits to calculate market sizes and market shares. To be consistent, total deposits are reported as 100 percent of bank deposits plus 50 percent of thrift deposits. 12. Bank regulators also review bank mergers affecting local rural markets for their.competitive effects. 13. Geographic boundaries of RMAs are given in Rand McNally's Commercial Atlas and Marketing Guide. Rand McNally states that there are two basic criteria which determine inclusion within an RMA. In general, an area must have at least 70 people per square mile, and at least 20 percent of the labor force must commute to the central urban area of the RMA. RMAs have been defined for all areas with a population of at least 50,000 and selected areas of less than 50,000. Tables 2a and 2b show that net increases in DIs were relatively rare; only 8 out of 65 urban markets (12.3 percent) showed a net increase in the number of DIs between 1982 and 1992. Data presented in Section III suggested that net decreases in the number of DIs in the Twelfth District were the result of very numerous mergers unmatched by a significant number of new DI charters. It is possible that mergers accounted for the elimination of fe\ver DIs in local markets than in the T\velftu~ District as a whole. This is because, unless the local markets of the DIs overlap, a merger or acquisition will not reduce the number of DIs, it will only change the target DI's name. In addition, the number of DIs in local markets can increase either through new charters or branching from outside of the market. However, the preponderance of markets with net decreases in DIs despite these factors suggests that many mergers may have been between DIs that operated in the same local urban market and that de novo branching into new local markets by established DIs may have been relatively uncommon. As shown in Section II, a decrease in the number of firms need not necessarily increase concentration if the inequality effect decreases. However, the majority of urban banking markets in the Twelfth District also did experience an increase in concentration and a presumed decrease in competitiveness between 1982 and 1992. Concentration increased in 43 markets (66.2 percent) and decreased in 22 (33.8 percent). Overall, average concentration in these 65 urban markets increased between 1982 and 1992, from an HHI of 1,643 to 1,747. Section III showed that average concentration also increased at the District level, but both 1982 and 1992 HHIs were much lower than in local urban markets. This is because DIs tend to operate in geographically restricted areas, so market shares are diluted in moving from the local to the District level, and concentration falls. The increase in average concentration also can be seen in Figure 1. Figure 1 shows that the number of markets in the second and third highest concentration categories increased, while the number of markets in the two lowest concentration categories decreased. The number of markets in the very highest concentration category stayed the same. However, note that the pattern of the distribution has remained roughly the same, with the largest number of markets having HHIs ranging from 1,200 to 1,499 in 1982 and 1992. Relatively few banking markets went from being "unconcentrated" or "moderately concentrated" in 1982 to "highly concentrated" in 1992. Anchorage, Honolulu, Hilo, Provo, Bellingham, Portland, Porterville, Eureka, an.d Tucson are the nine urban banking markets in which concentration went from below 1,800 in 1982 to at least LADERMAN / STRUCTURE OF URBAN BANKING MARKETS IN THE WEST 29 TABLE2A HHI, DIs, AND DEPOSITS IN TWELFTH DISTRICT RMAs WITH HHI INCREASES, RANKED BY CHANGE IN HHI RMA Anchorage, AK Honolulu, HI Hi],., , ......... HT .................. Provo, liT Santa Barbara, CA Bellingham, WA Portland, OR Porterville, CA Nogales, AZ Oxnard, CA Pocatello, ID Logan, UT Riverside, CA Fairbanks, AK Fresno, CA Eugene, OR Bakersfield, CA Saiem, OR Oceanside, CA Merced, CA Boise, ID Eureka, CA San Diego, CA Chico, CA Bremerton, WA Calexico, CA Corvallis, OR Modesto, CA Salt Lake City, UT Medford, OR Los Angeles, CA Palm Springs, CA Ogden, UT Longview, WA Lancaster, CA Monterey, CA Nampa, ID Phoenix, AZ Visalia, CA Davis, CA Hemet, CA Tucson, AZ Olympia, WA a 1992 level minus 1982 level HHI CHANGEa 1992 HHI 1,463 942 2,786 2,633 2,579 2,043 1,544 1,958 1,959 1,905 4,360 1,334 2,701 2,587 1,698 2,382 1,796 1,643 1,567 1,527 1,296 1,714 2,727 1,986 1,084 1,753 1,291 3,697 1,476 1,178 1,518 1,645 935 1,294 1,551 1,427 1,621 1,459 1,950 1,970 1,349 1,760 892 1,802 1,085 860 828 736 707 538 493 461 432 385 377 359 337 307 306 287 284 272 267 265 262 256 244 241 222 215 207 192 184 175 170 168 143 120 120 94 86 74 55 39 24 23 1992 DIs 6 14 9 12 18 12 25 7 3 25 6 6 45 5 22 15 16 12 19 10 10 9 67 9 15 5 10 19 26 13 247 25 12 10 11 15 7 37 15 8 19 13 16 DI CHANGEa -6 -2 -3 -8 -7 -4 -25 -4 -1 -6 -2 -4 6 -3 -7 -9 -6 -7 -6 -7 -3 -3 -4 -4 -4 0 -6 -9 -9 -6 -1 -10 -7 -6 -1 -4 -2 12 -2 -1 -4 0 2 1992 DEPOSITS (in thousands) 1,846,818 13,471,008 610,082 1,087,577 2,603,852 975,536 12,627,241 409,257 396,583 3,078,416 343,774 367,505 6,065,474 404,289 4,285,529 1,560,167 2,216,097 1,520,888 1,238,270 616,317 2,002,730 856,112 20,296,121 755,132 789,754 264,730 682,661 2,003,905 5,705,748 946,818 142,715,994 1,899,900 1,035,804 407,076 876,243 1,428,103 509,264 20,789,450 872,042 418,639 1,267,005 4,759,219 797,127 30 FRBSF ECONOMIC REvIEW 1995, NUMBER 1 TABLE 2B HHI, DIs, AND DEPOSITS IN TWELFfH DISTRICT RMAs WITH HHI DECREASES, RANKED BY ABSOLUTE VALUE OF CHANGE IN HHI RMA Stockton, CA Watsonviiie, CA Santa Cruz, CA San Francisco--Oakland, CA Yuba City, CA Las Vegas, NV Salinas, CA Yuma, AZ Redding, CA Idaho Falls, ID Fairfield, CA Santa Rosa, CA Santa Maria, CA Napa, CA Lewiston, ID Sacramento, CA Reno, NV Lompoc, CA Yakima, WA Seattle, WA Pasco-Kennewick-Richland, WA Spokane, WA a HHICHANGEa 1992 HHI 1992 Dis DICHANGEa -2,168 -681 -663 -638 -462 -379 -330 -329 -279 -256 -216 -213 -203 -125 -115 -106 -104 -69 -59 -43 -14 -13 1,217 1,622 1,288 1,424 1,408 1,822 1,252 1,866 1,624 1,805 1,495 878 1,492 1,296 1,377 1,241 2,392 1,977 1,388 1,589 1,894 1,722 22 -3 0 -6 -6 -1 1 -3 1 -3 0 -2 2 -1 -1 -1 -7 2 1 -3 -11 -5 -3 11 15 128 13 15 14 8 13 9 14 29 14 15 11 45 13 8 10 61 11 13 1992 DEPOSITS (in thousands) 2,637,372 589,672 1,555,037 87,220,301 773,151 6,763,318 1,319,883 575,167 1,151,968 602,044 850,920 2,525,034 904,191 951,443 428,995 10,152,322 2,195,682 295,032 942,979 25,171,338 747,770 2,561,116 1992 level minus 1982 level. 1,800 in 1992. In Alaska, the largest bank's acquisitions of several ofthe mid-sized banks in the state were allowed due to consideration of the acquired banks' poor financial conditions, leading to the inclusion of Anchorage in the above list. There are at least two possible reasons for the increases in the other markets. Frrst, the dynamics of competition may have caused shifts in market shares that would have increased concentration even in the absence of mergers. Second, the breach of the 1,800 level may be the result of the cumulative effect of multiple mergers, each of which passed the regulatory screen when considered on its own. 14 Similar reasons may have played a role in the Nogales, Pocatello, Logan, Fairbanks, Boise, and Calexico markets. All of these markets were already highly concentrated 14. For example, the market may start with an HHI of 1,650 and two separate mergers may be approved at different times, each of which increases the HHI by 100 points. in 1982 and saw cumulative changes of at least 200 points over the following ten years. Note that 7 of the 9 banking markets that went from being unconcentrated or moderately concentrated in 1982 to highly concentrated in 1992 rank in the top 8 banking markets in Table 2a in terms of increase in concentration. Accordingly, on average, the change in the HHI, at 680, for these 9 "crossover" markets, was considerably higher than the average change in the HHI of 203 for the other 36 markets that were unconcentrated or moderately concentrated in 1982. 15 However, it is also true that the crossover markets were, on average, more concentrated to begin with than the 36 noncrossover. markets. The average HHI in 1982 in the crossover markets was 1,504 and in the other 36 unconcentrated or moderately concentrated markets it was 1,298. 15. These 36 markets include 10 in which concentration decreased between 1982 and 1992. LADERMAN / STRUCTURE OF URBAN BANKING MARKETS IN THE WEST FIGURE 1 HHI IN TWELFTH DISTRICT RMAs NUMBER OF MARKETS 25 • 20 1 1982 ~ 1992 15 10 5 o < 1,200 1,2001,499 1,5001,799 1,8002,099 2 2,100 HHI For the sample as a whole, however, there appears to be a negative correlation between initial concentration and change in concentration. In the 43 markets in which concentration increased, the average HHI in 1982 was 1,517, whereas, in the 22 markets in which concentration decreased, the average HHI in 1982 was 1,888. In addition, with an average HHI increase of 331 and an average HHI decrease of 339, the absolute values of the average changes for the increasing and decreasing concentration groups were about equal to each other and about equal to the difference between the groups' initial average concentration levels. As a result, on average, the group with concentration increases ended up with about the same level of concentration as the initially high concentration group had in 1982, and the group with concentration decreases ended up with about the same level of concentration as the initially low concentration group had in 1982. Given the apparent tendency for concentration to increase in relatively unconcentrated markets and decrease in relatively concentrated markets, I tested whether there was in fact any statistical correlation. Using the urban banking market sample, I regressed the change in the HHI on a constant and the initial level of the HHI, using ordinary least squares. The coefficient on the 1982 HHI was indeed negative and highly statistically significant. This very simple fitted model indicated that, over a lO-year period, concentration increased about 596 points minus 29.9 percent of the initial HHI. This means that, according to the 31 model, markets in which the HHI is below about 1,985 tend to increase in concentration and markets in which the HHI is above that point tend to decrease in concentration. However, the model is misleading in that it specifies that the higher the initial concentration, the higher the terminal concentration. 16 In other words, according to the model, although concentration will fall in the more concentrated markets and rise in the less concentrated markets, the ordering of markets by HHI will not change. Tne flip in average concentration between the initially low concentration group and the initially high concentration group suggests that this is not necessarily the case. Concentration in a given market may fluctuate within a band, tending to increase up to the ceiling of the band ifit hits the floor of the band and tending to decrease down to the floor of the band if it hits the ceiling, thereby changing the concentration ordering of markets over time. The increase in the overall average indicates that any such band may have shifted up between 1982 and 1992. The model's specification of a decrease in concentration in the more concentrated markets along with an increase in concentration in the less concentrated markets and no change in the concentration ordering also erroneously suggests a decrease in the dispersion of concentration. In fact, the standard deviation of the HHI across urban banking markets barely changed between 1982 and 1992, increasing from 598 to 604. The negative correlation between initial concentration and the change in concentration partially may be a consequence of the application of the DOl's bank merger policy to individual mergers. Under a strict application of the policy, the farther below 1,600 is a pre-merger HHI, the larger an increase in the HHI will be permitted. This also is true for pre-merger HHIs between 1,600 and 1,800, but, here, 200 is the maximum change in the HHI allowed. (Under a strict application of the bank merger policy, for pre-merger HHIs of at least 1,800, there is no negative correlation between the level of the HHI and the permissible change in the HHI.) It may also be the case that the supracompetitive profits presumably found in very concentrated markets attract entry into those markets that helps to reduce concentration and restore competition. Underlying Causes for Increases and Decreases in Concentration Table 3 fills in the cells of Table 1 with the identities of urban markets in each category.17 Table 3 thereby shows 16. The 1992 HHI is approximately 0.7 times the 1982 HHI plus 596, according to the model. 17. There are 37 markets represented in Panel A of Table 3. These are the markets with initial variance ofmarket shares above the critical value 32 FRBSF ECONOMIC REVIEW 1995, NUMBER 1 TABLE 3 HHI CHANGES AND THEIR CAUSES IN TWELFfH DISTRICT URBAN MARKETS N Increased V Increased + + Riverside Tucson NV Increased V Decreased + + Portland Fresno Bakersfield Boise Modesto Los Angeles Honolulu Oxnard Eugene Oceanside San Diego Salt Lake City Palm Springs n.a. n.a. Reno Phoenix n.a. n.a. V Increased NDecreased No Change in N + Medford Ogden Monterey Visalia NV Decreased Sacramento Seattle Pasco/Kennewick/Richland Spokane V Decreased Watsonville Las Vegas Santa Rosa PANEL B: Vo < (1O,OOOINoN1) No Change in N N Increased V Increased NDecreased + + Lompoc Olympia Calexico NV Increased + n.a. n.a. + n.a. Portland Logan Merced Chico Longview NV Decreased Pocatello Salem Eureka Corvallis Hemet + V Decreased Yuma NOTE: NV = inequality effect V = variance N = number of firms Hilo Santa Barbara Bremerton Anchorage Provo Bellingham Lancaster n.a. V Decreased V Increased Santa Cruz Yuba City Salinas Fairfield Napa Stockton San Francisco/ Oakland Redding Santa Maria Idaho Falls Nogales Fairbanks Nampa Davis Lewiston Yakima LADERMAN / STRUCTURE OF URBAN BANKING MARKETS IN THE WEST underlying causes for increases and decreases in concentration in each market. For example, concentration increased in the Boise banking market because the number of DIs decreased and variance increased enough that the inequality effect also increased. On the other hand, concentration increased in the Riverside banking market because, even though the number of DIs increased, the variance of market shares increased, and initial variance was above the critical value of 10,000 divided by the product of the 1982 and 1992 number of DIs. Equivalently, concentration increased in the Riverside market because, even though the number of DIs increased, the variance of market shares also increased enough that the increase in the inequality effect outweighed the negative effect that an increase in DIs by itself would have had on concentration. The experiences of the largest markets, those with over $10 billion in deposits in 1992, varied somewhat. These markets are Honolulu, Portland, San Diego, Los Angeles, Phoenix, Sacramento, Seattle, and San FranciscoOakland. In the Honolulu, Portland, San Diego, and Los Angeles markets, concentration increased because the number of DIs decreased and the variance of market shares increased enough that the inequality effect also increased. In the Phoenix market, even though the number of DIs increased and· the variance of market shares decreased, concentration increased because the inequality effect increased sufficiently. Note that the increase in the inequality effect in the Phoenix market was due solely to an increase in the number of DIs. The experience in the Sacramento and Seattle markets was the opposite of that in the Phoenix market. In these markets, even though the number of DIs decreased and the variance of market shares increased, concentration decreased because the inequality effect decreased sufficiently. The decreases in the inequality effects in the Sacramento and Seattle markets were due solely to decreases in the number of DIs. In the San FranciscoOakland market, even though the number of DIs decreased, concentration decreased because variance also decreased, so that the decrease in the inequality effect outweighed the concentrating effect that a decrease in the number of DIs has by itself. Note that, because initial variance was above the critical value in the San FranciscoOakland market, a decrease in the number of DIs had to decrease concentration if variance either did not change or decreased. 18 derived in Section II. There are 28 markets in Panel B of Table 3, with initial variance below the critical value. 18. Note, however, that, given the initial number of firms, the larger the decrease in the number of firms, the higher initial variance must be to exceed the critical value. 33 The counts of markets in each cell also suggest general conclusions regarding underlying causes for changes in competitiveness and efficiency, as measured by concentration, in Twelfth District urban markets. In 38 of the 53 markets in which the number of DIs decreased, concentration increased. In twenty of these markets, the change in the inequality effect reinforced the effect of the decline in the number of competitors. In the others, the change in the inequality effect partially mitigated the effect of the decrease in the number of DIs, but not enough to outweigh the concentrating effect that a decrease in the number of firms has if the inequality effect is held constant. There were 15 markets that became less concentrated despite a decline in the number of DIs. Concentration decreased in these markets because the decrease in the inequality effect outweighed the concentrating effect of a decrease in the number of firms, holding the inequality effect constant. In markets overall, increases in the variance of market share were more common than decreases (43 increases versus 22 decreases). However, due to the preponderance of markets in which the number of DIs decreased, decreases in the inequality effect were more common than increases (38 decreases versus 27 increases). Average sizes of increases and decreases in concentration depended on whether the direction of change in the number of DIs, the variance of shares, and the inequality effect worked in the same direction or not. In the 33 markets in which, given only the direction of change of these factors, concentration had to increase, the average increase in the HHI was 382. In the 14 markets in which concentration had to fall, the average decrease in the HHI was 496. 19 For the 18 markets in which the effects worked in opposite directions, the average absolute change in the HHI was 118. For markets in which the inequality effect and change in the number of DIs worked in the same direction, one can calculate the proportion of the change in concentration that was due to each factor. For the 3 such markets in which concentration decreased, on average 73.3 percent of the 19. The Stockton, California market, which showed an HHI decrease of 2,168, may be considered to be an outlier. The large decrease in concentration in the Stockton market primarily was due to an outflow of deposits from the largest DI, a thrift. This thrift had been paying above market interest rates to attract deposits and held 56 percent of the deposits in the market in 1982. When it encountered financial trouble and stopped paying high rates, it lost deposits to other DIs in the market, greatly reducing the overall variance of market shares. If the Stockton market is excluded, the average decline in the HHI in markets in which concentration had to decrease was 367. 34 FRBSF ECONOMIC REVIEW 1995, NUMBER 1 change in concentration was due to the decrease of the inequality effect. For the 20 such markets in which concentration increased, on average 56 percent of the change in concentration was due to the increase of the inequality effect. 20 These percentages suggest that when the change in the inequality effect and in the number of DIs (holding the inequality effect constant) work together, the former is somewhat more important than the latter. V. CONCLUSION Both the Twelfth District as a whole and local urban banking markets in the District saw a widespread reduction in the number of DIs between 1982 and 1992. Nearly 82 percent of the 65 urban banking markets in the District saw a net decrease in the number of DIs. Although this trend need not necessarily have been accompanied by an increase in concentration, in most markets it was. As a result, concentration in the Twelfth District overall increased between 1982 and 1992. Concentration also increased in approximately two-thirds of the 65 urban banking markets in the Twelfth District. On the other hand, there is some evidence that if a market becomes concentrated enough, concentration will start to fall, thereby helping to strengthen competition and productive efficiency. The preponderance of urban banking markets in which the number of DIs decreased likely was a consequence of numerous within-market mergers that were unmatched by significant numbers of new charters. District level data also suggest that a large proportion ofthe net decrease in DIs was accounted for by a net decrease in the number ofthrifts. This likely also played a role in local urban markets. Given a slowdown in the disappearance of thrifts, the decreasing trend in the number of DIs should abate somewhat. Shifts in market shares can reduce the effect of the inequality of shares on measured concentration. If this inequality effect decreases sufficiently, it can overcome the concentrating effect that a decrease in the number of firms has if the inequality effect is held constant. However, less than a third of the markets in which the number of DIs decreased showed decreases in concentration. This suggests that regulatory review of bank and thrift mergers and acquisitions has been and will continue to be important in assuring the competitiveness of banking markets. 20. Markets counted exclude those in which the number of DIs did not change. REFERENCES Commercial Atlas and Marketing Guide. 1992. 123rd edition. Skokie, Illinois: Rand McNally. Holdsworth, David G. 1993. "Is Consolidation Compatible with Competition? The New York and New Jersey Experience." Federal Reserve Bank of New York Research Paper #9306. Rhoades, Stephen A. 1993. "The Herfindahl-Hirschman Index." Federal Reserve Bulletin (March) pp.l88-189. Scherer, EM. 1980. Industrial Market Structure and Economic Performance. 2nd edition. Boston: Houghton Mifflin.