The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Vol. 31, No. 3 ECONOMIC REVIEW 1995 Quarter 3 f Inflation, Unemployment, and Poverty Revisited 2 m by Elizabeth T. Powers Scale Economies and Technological Change in Federal Reserve ACH Payment Processing by Paul W. Bauer and Diana Hancock FEDERAL RESERVE BANK OF CLEVELAND 14 1 E C O N O M I C R E V I E W 1995 Quarter 3 Vol. 31, No. 3 Inflation, Unemployment, and Poverty Revisited 2 by Elizabeth T. Powers Most of the research that uses income to measure economic well-being shows that while unemployment has a strong positive effect on poverty rates, inflation has very little effect. This paper considers the impact of inflation and unemployment on poverty, using a poverty rate based on goods and services actually consumed, rather than on income. The find ings suggest that increases in unemployment are associated with increases in both the consumption poverty rate and the conventional income poverty rate. However, inflation seems to have a robust and relatively large positive influence on consumption poverty, indicating that inflation may harm the poor more than was previously thought. Scale Economies and Technological Change in Federal Reserve ACH Payment Processing 14 by Paul W. Bauer and Diana Hancock Since 1979, the cost to the Federal Reserve of processing an automated clearinghouse (ACH) transaction has fallen dramatically. The authors of this study find that three factors— scale economies, technological change, and lower input prices— each contributed significantly to this price decline. Their results also show that substantial scale economies could still be achieved in ACH payments processing. This research should be of broad interest to economists because the data provide a rare, detailed glimpse into the workings of a service industry. Economic Review is published quarterly by the Research Depart ment of the Federal Reserve Bank of Cleveland. Copies of the Review are available through our Corporate Communications & Community Affairs Department. Call 1-800-543-3489 (OH, PA, WV) or 216-579-2001, then im mediately key in 1-5-3 on your touch-tone phone to reach the publication request option. If you prefer to fax your order, the num ber is 216-579-2477. Coordinating Economist: Jagadeesh Gokhale Advisory Board: Charles T. Carlstrom Joseph G. Haubrich Peter Rupert Editors: Tess Ferg Michele Lachman Robin Ratliff Design: Michael Galka Typography: Liz Hanna Opinions stated in Economic Review are those of the authors and not necessarily those of the Fed eral Reserve Bank of Cleveland or of the Board of Governors of the Federal Reserve System. Material may be reprinted pro vided that the source is credited. Please send copies of reprinted material to the editors. ISSN 0013-0281 2 Inflation, Unemployment, and Poverty Revisited by Elizabeth T. Powers Elizabeth T. Powers is an econo mist at the Federal Reserve Bank of Cleveland. The author thanks David Altig, Alan Auerbach, Louise Scheiner, Daniel Slesnick, and the editorial board ot the Eco nomic Review for helpful com ments, and Kristin Roberts for valuable research assistance. Introduction A small but influential body of literature has attempted to estimate the effect of selected macroeconomic variables on poverty.1 Such exercises may serve several purposes. For example, general knowledge of predictable empirical relationships among these variables might aid fiscal planning. However, most of this work has been motivated by “the frequent out cries against inflation on the grounds of its adverse effects on the distribution of income.”2 This literature consistently finds that inflation has a relatively minor impact on the incidence of poverty and on the well-being of poor and near-poor households. Because most econo mists working in this area assume that there is a direct trade-off between inflation and unem ployment, controllable by the policymaker, the critical comparison is between the effects of the inflation and unemployment rates on poverty. H 1 For example, while it is not the focus here, aggregate economic growth is a frequently used macroeconomic indicator variable in this litera ture. See Powers (1995) for a discussion. I 2 See Blinder and Esaki (1978), p. 604. This paper considers the relationship be tween these macroeconomic variables and an alternative poverty measure that is based on consumption rather than income. Otherwise, I follow the methodology of the existing literature closely. My research findings suggest that changes in the unemployment rate are impor tant in explaining variation in both the conven tional income poverty rate and a consumptionbased poverty rate (which I call the JS poverty rate, after work by Jorgenson and Slesnick [1987, 1990] and Slesnick [19931). However, in sharp contrast to previous findings that inflation has very little effect on income poverty , I find a robust and relatively large positive relationship between inflation and the consumption poverty rate. Thus, my findings suggest that inflation may have a more adverse effect on poverty than was previously thought. Before explaining the methodology and find ings, it is important to note that there are several possible avenues for improving on the existing literature. Perhaps most seriously, the relation ship between inflation and unemployment, long a subject of intense debate, is not modeled. Typically, aggregate indicators of poverty such as the share of all income received by the 20 percent (quintile) of households reporting the lowest income, or the poverty rate (the percent of the population living in households with income below a given level), are simply re gressed on measures of unemployment and inflation. Inflation and unemployment rates are treated as if they have no influence on each other, or are not both partly determined by some common factor. This is at odds with most theoretical treatments of the macroeconomy, and ignoring the existence of these relationships can result in unreliable estimates. Use of the quintile share of income as a pov erty indicator can also be misleading. In many cases, this variable is not informative about changes in the welfare of the poor. For exam ple, suppose that households in the top income quintile are taxed and the proceeds destroyed. By definition, the total income share of the bot tom group must rise, yet it is obvious that this latter group is not better off in any substantive way. For similar reasons, empirical estimates of the influences of inflation and unemployment on quintile shares are not easily interpreted. In flation or unemployment may harm low-income groups absolutely, even while their effects on quintile shares are positive or negative.3 Finally, except for the work of Cutler and Katz (1991), this literature has developed under the assumption that income poverty concepts adequately measure economic well-being. In the past, this has been a matter of necessity, because income data were the most compre hensively and consistently collected. However, economic theory suggests that the goods and services actually consumed by a family or indi vidual are a better measure of their well-being (the economist’s ideal measure being utility). Poverty measures based on income and con sumption are expected to differ because, in principle, money income and consumption can differ substantially.4 This means that who is classified as poor can vary across the two meas ures. Further, the predominant economic model of consumption argues that households attempt to protect their standard of living from short term income swings. This implies more yearto-year variation in household income than in consumption. Hence, the income poverty count should also include more families who are transitorily poor, while consumption poverty should include more families who view their status as persistent. For all these reasons, rela tionships found to hold with respect to poverty measured on an income basis may not be ro bust with respect to poverty measured on a consumption basis. Because of the difficulties in interpreting the quintile-share measures of well-being, I focus exclusively on the poverty rate.5 However, the poverty rate has some severe limitations of its own. After all, it is merely a head count of those below a particular threshold, and changes in macroeconomic conditions can dramatically affect the well-being of the poor without chang ing the actual head count at all. Therefore, it is important to remember that the poverty rate portrays only a single (albeit important) feature of the nation’s poverty situation. While the modeling of the macroeconomy in previous work is obviously open to question, there is so little agreement on the proper model that such an approach is unappealing. Instead, I accept the premises on which the previous lit erature rests, and ask whether these findings are robust with respect to the poverty concept employed. Thus, this paper is best interpreted as a sensitivity analysis of the previous findings vis-a-vis inflation, unemployment, and the poverty rate. The paper’s first section discusses and inter prets the findings of the previous literature. Section II traces the history of the official meas ure of income poverty and considers its flaws. The development of the alternative historical series of consumption poverty rates presented in Slesnick (1993), and the differences between it and the conventional poverty series, are dis cussed in section III. Section IV revisits the issue of inflation, unemployment, and poverty using alternative poverty and inflation meas ures. Section V concludes. I. Unemployment, Inflation, and the Conventional Poverty Rate In this section, I discuss and update the previ ous literature’s findings on inflation, unemploy ment, and income poverty. To interpret the ■ 3 In fact, it is easy to construct a model in which the impact of inflation is consistently positive or negative on all incomes, but the rela tionship between inflation and any quintile’s share, and even the ratio of low to high shares, is nonmonotonic. ■ 4 In theory, low-income households could also be drawing on savings, borrowing against future income, receiving gifts or government transfers of goods and services, or even getting income from the under ground economy. Of course, whether they actually do so is an empirical question. ■ 5 While the poverty rate has its own limitations, at least its pre dicted relationship with the variables of interest here is unambiguous. findings, however, it is important to consider the microfoundations of income poverty and to understand how changes in macroeconomic variables are transmitted into changes in poverty rates. How might higher overall unem ployment affect the number of persons living in poverty? The majority of families rely on labormarket earnings for most of their income, so episodes of unemployment may result in large income declines. It is also well known that unemployment in cyclical downturns is dispro portionately borne by people whose earnings are low to begin with— those whose incomes are most vulnerable to slipping below the poverty level.6 These factors are expected to produce a strongly positive relationship be tween unemployment and poverty rates. How ever, there are other, potentially mitigating, fac tors. Some have theorized that the pattern of wages over the business cycle could be pro cyclical,7 and that dependency on government transfer payments might also lessen the poverty rate’s sensitivity to unemployment by reducing the role of earned income. Dependence on unindexed income is the obvious channel through which inflation might affect income poverty rates. Households that rely on nominally fixed payments for a substan tial portion of their income could be driven into poverty by inflation; this implies a positive rela tionship between inflation and poverty rates. The primary sources of nominally fixed income are means-tested transfer payments (Aid to Families with Dependent Children [AFDC] and states’ General Assistance programs being the only significant unindexed cash-transfer pro grams) and the minimum wage.8 It is also pos sible that employers exercise temporary market power in inflationary periods, allowing real wages to fall in the short run. Finally, as the next section discusses, the poverty line was probably overindexed for inflation throughout the 1970s and 1980s, implying that some por tion of poverty-rate increases may be explained by increased inflation itself. The primary studies on inflation, unemploy ment, and the size distribution of income in the United States are those of Blinder and Esaki (1978), Blank and Blinder (1986), Blank (1993), Cutler and Katz (1991), and Mocan (1995). Ex cept for Blinder and Esaki (who estimate only income shares), all of these studies estimate straightforward empirical relationships between poverty rates and macroeconomic variables.9 Blank and Blinder (1986) first examined the relationship between unemployment, inflation, and official income poverty rates for families and persons. Their regression findings indicate that inflation and unemployment rates were both positively related to the percent of all per sons living in poverty during the 1959-1983 period. However, while inflation was associated with an increase in the steady-state poverty rate, this effect was only one-seventh the mag nitude of the poverty-increasing effect of a rise in the unemployment rate. This led Blank and Blinder to conclude that while both unemploy ment and inflation worsen poverty, the empiri cal evidence supports their belief that “unem ployment, not inflation, is the crudest tax.” Blank (1993) also found a significantly positive relationship between inflation and poverty rates. In contrast, Cutler and Katz (1991) and Mocan (1995) reported a relatively small nega tive relationship between inflation and poverty. A strong, robust, positive relationship between poverty and unemployment has been consis tently observed. Because of revisions to data series as well as the availability of new data since the original studies appeared, I have updated some repre sentative findings in the literature using the poverty rate for persons, as computed by the Census Bureau from 1959 to 1992 (table 1). The specification in the first column includes an intercept term, an inflation measure (the growth rate of the Consumer Price Index for all urban consumers, or CPI-U), the unemployment rate for prime-age males, and additional explanatory variables, including the ratio of the poverty level for a family of four to mean household income, and a trend for the years after 1983. In the sec ond column, the one-period lag of the poverty rate is added to the specification as a crude con trol for any dynamic features of the evolution of poverty.10 The unemployment rate for males ■ 6 While it is probity safe to assume that tamilies starting out nearest the poverty line are most vulnerable to crossing it, there is also substantial income mobility from year to year in the U.S. economy. It is possible that some people whose incomes put them well above the poverty line one year might find themselves below it the next. ■ 7 The evidence on this matter is inconclusive. ■ 8 It is doubtful that these income sources exert an important influ ence on the poverty rate. Very few families of any kind contain a minimum wage earner (see, for example, Horrigan and Mincy (19931). And, while real AFDC benefits have been declining over the past 20 years, the effect on per capita benefits has largely been mitigated by dedintng household sizes 9 A number of studies apply this methodology to foreign economies, a recent example being Yoshino (1993) on Japan. Minarik (1979) used an alternative microsimulation approach to examine the effect of inflation alone on the size distribution of income. ■ 10 The specifications reported in the first two columns are similar to those presented in Cutler and Katz (1991). ■ 5 TABLE 1 Regression Findings for Income Poverty, 1959-1992 Dependent Variable: Income Poverty Rate for Persons Explanatory Variables -10.443 (1.29)b Constant Poverty line/ mean income -6.65a (1.09) Inflation (growth in CPI-U) 0.635a (0.029) -0 .1 14a (0.043) 0.2893 (0.058) 0.065c (0.039) Prime-age-male unemployment rate 0.433a (0.068) 0.323a (0.046) Post-1983 trend 0.3383 (0.054) 0.1992 (0.042) -5.10 (4.185) 0.366a (0.111) 0.081c (0.049) 0.584a (0.224) Lagged-income 0.586a poverty rate (0.090) Post-1982 dummy (1983-1992 = 1) Government transfers to persons/GNP Interactions with Post-1982 Dummy 0.371a (0.119) -3.41 (2.39) -0.278 (0.237) Prime-age-male unemployment rate -0.748a (0.190) Government transfers to persons/GNP 0.787a (0.247) Inflation (growth in CPI-U) -0.039 (0.149) 96.8% Adjusted R2 34 Number of observations 98.5% 33 98.6% 33 a. Significantly different from zero at the 5 percent level or greater. b. Standard errors are in parentheses. c. Significantly different from zero at the 10 percent level or greater. SOURCE: Author’s calculations. aged 25-54 is used to capture unemployment effects on poverty, since the total unemploy ment rate is influenced by demographic trends that may independently affect the income poverty rate. The ratio of the poverty line to mean household income is intended to control for the shape of the income distribution near the poverty line (see Danziger and Gottschalk [1986]). Finally, the post-1983 trend attempts to account for that era’s unusually and persis tently high poverty rate (Cutler and Katz [1991], Blank [1993]). In the first two columns, the unemployment rate shows a strong positive effect on the in come poverty rate.11 An increase of one per centage point in the prime-age-male unem ployment rate raises the poverty rate by an estimated 0.3 to 0.4 percentage point. Accord ing to the first column, periods of high inflation are associated with poverty-rate reductions. An increase of one percentage'point in the infla tion rate leads to a reduction of 0.1 percentage point in the poverty rate. However, the findings with respect to inflation are sensitive to specifi cation; the findings reported in the second col umn suggest that inflation has a (weakly) posi tive effect on the income poverty rate. The final specification, reported in the last column, is similar to those in Blank (1993) and Blank and Blinder (1986). In addition to the previous variables, Blank includes a measure of government policy (government transfers to persons divided by GNP) and tests for structural change in the relationship between unemploy ment, policy, and poverty after 1982.1 have added a term to test for a structural change in the inflation effect as well. Unemployment has the strongest effect in this specification, while inflation has only a weakly positive impact. All of the macroeconomic variables appear to have perverse effects in the post-1982 world, as noted by Blank. Recently, Mocan (1995) has presented a more elaborate econometric treatment of the relationships between unemployment, inflation, and poverty. He specifies poverty rates as a function of unemployment, inflation, and real wages, and uses a “flexible” model of the trend in the poverty rate. The problem is that the deterministic trends previously used in this liter ature may be inappropriate if the trend in the poverty rate is subject to stochastic disturbances. This issue is important because proper detrend ing of the data is critical for reliable parameter estimates. Mocan also decomposes unemploy ment into its short- and long-run components and inflation into its anticipated and unantici pated components, and analyzes black and white poverty rates separately.12 While Mocan finds that cyclical unemployment has almost no effect on income poverty, long-run (structural) unemployment has a significantly positive ef fect. He also finds that both expected and unex pected inflation significantly redu ce poverty, ■ 11 It should be noted that to preserve comparability with previous studies, I do not correct for the obvious autocorrelation in all of the specifi cations in table 1. However, corrected estimates (which are not reported) are qualitatively similar. ■ 12 Blank and Blinder (1986) also decomposed inflation, but found no significant differences between unanticipated and anticipated inflation effects. 6 with the former having the larger impact. The negative effect of inflation on the person poverty rates for blacks and whites is about one-third of Mocan’s estimated poverty reductions from a decrease in structural unemployment. To summarize the literature’s findings, un employment is consistently estimated to have a strong positive effect on the income poverty rate, suggesting that joblessness is responsible for pushing many households’ incomes below the poverty level. This finding is quite robust with respect to various empirical specifications. While the estimated effect of inflation is very sensitive with respect to specification, it seems to have at most a small positive impact on the poverty rate, and may even be associated with poverty-rate declines. Unfortunately, these findings are developed in the context of a poorly specified measure of poverty. A consumption-oriented approach to poverty suggests that the important factors are the total resources available to a family over long periods, and the family’s ability to rearrange these resources over time. If consumption and income poverty rates turn out to be very differ ent, one expects that the findings vis-a-vis infla tion, unemployment, and poverty will also be very different— for two reasons. First, as I dis cuss below, the mechanisms by which unem ployment and inflation may be translated into consumption poverty are quite different from those influencing income poverty. This suggests that the relationship of macroeconomic vari ables to consumption poverty is potentially very different from their relationship to income poverty. Second, the income poor and the con sumption poor may be dissimilar groups of peo ple. (For example, they appear to vary in age and racial composition, according to Slesnick [1993].) Since the response to macroeconomic conditions is undoubtedly heterogeneous across the population, changing the type of house holds under consideration should also change the aggregate relationships. Of course, if income poverty is a close approximation of the underlying “true” con sumption poverty rate, these issues will be sig nificant only in theory, not in practice. In the next section, I review Slesnick’s (1993) calcula tion of consumption poverty.13 ■ 13 The material in section II is drawn primarily from Slesnick (1993) and Ruggles (1990). II. The Mismeasurement of Poverty A Brief History of the Poverty Line The official government poverty rate is the pro portion of the population whose pretax income falls below specified levels, called “poverty thresholds” or “poverty lines.” Today’s official poverty thresholds have their antecedents in the poverty lines developed for the Social Secu rity Administration by Orshansky (1988) in the early 1960s. Because budget studies from the 1950s found that the typical low-income family spent about one-third of its budget on food, Orshansky took the USDA’s Economy Food Plan (a nutritionally adequate but inexpensive collection of food items) and multiplied it by three to arrive at a level of total expenditures designated as the poverty line. Poverty thresholds were further refined for the heterogeneous nutritional requirements of families with different structures. Until 1981, a particular family’s threshold depended on family size, sex and age of the household’s head, num ber of related children under 18, and farm or nonfarm residence. Smaller families devote a rel atively smaller share of total expenditures to food; women have lower caloric requirements than men; children eat less than adults; and farm families can consume home-grown food. All of these considerations suggested lower multiplica tive factors, and hence lower poverty expendi ture thresholds, all else being equal.14 In 1981, calculations of differences due to sex of the household head and farm versus nonfarm resi dence were eliminated by legal challenges. Nominal thresholds must be adjusted over time to reflect declines in purchasing power. Prior to 1981, the nominal poverty line was increased by food-price inflation only. By ignoring other prices, these adjustments some times overstated, sometimes understated, the increase in total nominal expenditures required to maintain a constant standard of living. Since 1981, the CPI-U has been used to inflate the official poverty thresholds from their 1963 val ues to current dollars. ■ 14 Indeed, it is possible to differentiate along many more charac teristics, as suggested by Slesnick (1993). FI GURE 1 Alternative Inflation Rates, 1951-1989 Annual percentage change 14 12 10 8 6 4 2 0 -2 1951 1955 1959 1963 1967 1971 1975 1979 1983 1987 SOURCE: Author’s calculations. Problems with the Official Poverty Rate As Slesnick (1993) points out, the conceptual basis for the official poverty statistic, based on an expenditure concept, is fundamentally sound. However, several features of the poverty thresh olds are simplistic and may bias the measure ment of poverty. Foremost among these are benchmarking against food consumption and the inflation adjustment. Using family equiva lence scales based entirely on food needs will in some cases understate, and in other cases over state, efficiencies in the shared consumption of nonfood commodities. For example, a childless couple may need almost twice as much food as one person, but they will not need twice as many rooms in their apartment. Thus, multiply ing their Economy Food Plan figure by three may lead to a gross overstatement of their mini mal expenditure requirements. Because the food equivalence scale will understate efficiencies of shared consumption in other items, the direction of the total bias that results from relying solely on food shares is unpredictable a priori. Several obvious issues are raised by adjusting the poverty thresholds by a single inflation rate each year, and several problems are peculiar to the CPI-U. First, an increase in the general price represents the combined effect of increases (and/or decreases) over all prices, but all prices do not necessarily rise at the same rate. When, for example, inflation is concentrated in the price of necessities, the poor, who devote a greater fraction of total expenditures to these items, will be harmed more than others. This suggests that poverty thresholds should be ad justed by price indexes that are relatively more sensitive to rising prices of items consumed in tensively by the poor, rather than by the CPI-U, which reflects inflation based on expenditure patterns of the average family. Another potential problem of applying a single inflation measure to poverty thresholds is that expenditure patterns may adjust in ways that mitigate welfare losses from price changes. In theory, families can accommodate fairly sig nificant inflationary episodes by adjusting the types and relative quantities of goods they consume.15 For instance, when beef prices rise relative to chicken prices, consumers may sub stitute chicken for beef. These behavioral re sponses result in smaller declines in living standards than if expenditures remained frozen in their former patterns. Since the CPI-U is only infrequently reweighted for changes in expen diture patterns (and not of the poor, but of the average family), applying it to the poverty line overstates the increase in poverty thresholds required to approximate the same level of well-being. A final problem, peculiar to the CPI-U itself, is its treatment of housing. Before 1984, the housing component was set equal to the finan cial cost of housing, not the flow of housing services. Thus, periods of high mortgage rates are periods of overstated inflation in the CPI-U series. Figure 1 shows both the CPI-U and the alternative CPI-X1, which uses rental costs as a proxy for housing service prices. The CPI-U overstates inflation in the late 1960s and late 1970s, implying that poverty thresholds rose by more than the amount needed to maintain a constant standard of living, and overstating recent poverty rates. After 1984, the two price indexes are the same. With the exception of the housing error, the above factors make a relatively minor contribu tion to the mismeasurement of poverty (Slesnick [19931). The most serious divergence between theory and implementation is the use of pretax income, rather than expenditures, as the yardstick for poverty. This practice accounts for most of the mismeasurement of poverty. In the next section, I explore the construction of alternative consumption-based poverty rates and the biases introduced by the use of income- rather than consumption-based rates. ■ 15 That is, substitution as well as income effects are associated with price changes. 8 III. A ConsumptionBased Poverty Rate The accurate estimation of consumption-based poverty rates is a daunting task. Slesnick (1993) overcomes several obstacles to arrive at a series that addresses the many problems discussed in the previous section. His estimates are devel oped under the assumption that families act as life-cycle consumer units, saving and dissaving to smooth consumption over time.16 An im plicit assumption is that the fraction of “misers” in the population is small. Presumably, for most families, a consumption poverty classification reflects low resources rather than a preference for low consumption. Slesnick’s basic consumer data are from the Consumer Expenditure Surveys (CES) for 19601961, 1972, 1973, and 1980-1989.17 Measuring poverty on the basis of consumption, rather than income, is not a simple matter of compar ing CES expenditure data to the standard poverty thresholds. First of all, expenditure and consumption are not equivalent concepts. For example, contributions to retirement funds (including Social Security taxes), which the CES records as expenditures, are really savings, since they contribute directly to future living standards. Contributions or gifts to other house holds, while available, are not used by Slesnick, since a consistent treatment would greatly com plicate the modeling of consumption. As in computing official poverty status, Slesnick excludes in-kind transfers of housing subsidies and health care from his measure of consump tion, although conceptually they should be included. Finally, many goods are consumed over long periods and not immediately upon purchase. Expenditures for these “durable” goods may occur all at once or over a period of years (homes and cars are frequently paid for in this way). There is no reason to expect pay ment schemes to exactly match the flow of value from the consumption of these services. Instead, Slesnick imputes the rental equivalent (what one would be willing to pay to rent the identical item) for durables in each survey year. The JS equivalence scales used to adjust for differences across family types are more detailed than the official equivalence scales. They measure how expenditure patterns for all items (not just food) change when household composition changes. In contrast to the official rates, which are based solely on nutritional requirements, and which vary only according to size and age characteristics of families and individuals, the “JS equivalence scales ... vary over any set of demographic attributes that influence household expenditure patterns” (Slesnick [1993DSlesnick addresses many of the indexing problems associated with the conventional poverty rate. For any combination of price changes, he estimates the minimum nominal change in total expenditures necessary to main tain a constant standard of living, which amounts to a specific cost-of-living index for each household. The index is applied to the base-year poverty threshold (which has been converted to a consumption-equivalent basis). This general deflator accounts for several factors excluded by the CPI-U, including the fact that price changes affect families with different con sumption patterns differently and lead to substi tution of less expensive for more expensive commodities. However, Slesnick shows that these adjustments’ effects on measured poverty are quite small. The primary impact on poverty rates comes through the correction for the over statement of inflation in the CPI-U due to the mistreatment of owner-occupied housing costs. Figure 2 shows the official income poverty rate and the JS consumption poverty rate for 1959-1989. Both the levels and trends of the two rates are quite different. Except for a period in the late 1960s, the official poverty rate is higher than the JS rate. Both rates decline from 1961 to the beginning of the 1970s. However, they paint dramatically different pictures of recent poverty trends. Because the JS family equivalence scales set a relatively lower poverty threshold for female heads than do the official equivalence scales, JS poverty continues to trend downward over the 1970s, when the pro portion of female heads in the general popula tion was rising dramatically and pushing up the conventional poverty rate.18 While the official rate indicates a strong resurgence in poverty throughout the early 1980s and persistently high rates thereafter, the JS rate, after a sharp increase around the time of a recessionary trough in ■ 16 Since a brief discussion can convey only the major contribu tions, readers with a deeper interest in the methodology and implementa tion should consult Slesnick (1993) and the references therein It should also be noted that the application of the life-cycle model to low-resource households is controversial. ■ 17 Imputation methods involving auxiliary information from the Current Population Surveys are used to derive poverty rates for the years not covered by the CES. Given the available dafla, this is the best one can do. The imputation process probably errs on the side of making the con sumption poverty measure aid the official poverty measure more similar. 18 The primary reason for the lower JS thresholds for female headed households is that children in these families are on average younger than children in two-parent families and so consume less. ■ g FI GURE 2 Alternative Poverty Rates, 1959-1989 Percent ol persons in poverty SOURCES: Author’s calculations; and Slesnick (1993). 1980, shows continued progress in the war on poverty throughout the 1980s. Although he does not compute the degree of overlap between the income-poor and consumption-poor groups, evidence provided by Slesnick from the CES supports the notion that the officially poor group is dominated by those with only temporarily low incomes and fairly high consumption. For example, in a typi cal year, 40 percent of the income poor are homeowners (as are 60 percent of the general population); in contrast, only 17 percent of the consumption poor own their homes. Thus, a significant minority of the income poor receive substantial service flows from housing, while most of the consumption poor do not. The consumption poor also devote a larger share of total expenditures to necessities such as food (ranging from 31.6 percent to 37.3 percent over the 1961-1989 period) than do the income poor (22.2 percent to 28.1 percent). The life cycle model implies that dissavers view their low income as a transitory circumstance; in deed, Slesnick finds substantial dissaving occur ring among the income poor. While 59 percent to 76 percent of the income poor dissave over the CES surveys, only 21.7 percent to 36.4 per cent of the consumption poor dissave, suggest ing that the consumption poor view their lack of resources as a permanent condition. The divergence between the two poverty measures is expected to grow over time, since the poverty line is an absolute— not a rela tive— notion of well-being. When average income is fairly low, there are relatively more people whose “typical” annual income is near http://fraser.stlouisfed.org/ or below the poverty line. As average real Federal Reserve Bank of St. Louis income grows, as it has since I960, there are relatively fewer people whose typical income is below the poverty line. Thus, the income-poor population is increasingly dominated by people with extraordinarily bad income realizations— just the people for whom consumption does not equal income. IV. Inflation, Unemployment, and Poverty Revisited Before discussing the empirical approach and findings, it is useful to describe the ways in which inflation and unemployment might be expected to influence consumption poverty. Unemployment may affect consumption poverty rates in several ways. If the household is liquidity constrained (so that consumption is limited to current income), then a spell of un employment may result in both income and consumption poverty. When the household is not liquidity constrained, a spell of unemploy ment should be harmful to the extent that it decreases permanent, but not transitory, in come. For example, if earnings losses in reces sions are offset by increased opportunities in expansions, cyclical unemployment should not affect permanent income or consumption pov erty rates. However, if the labor market rewards continuity in employment, time out of the labor force may permanently reduce income, and hence consumption. Finally, periods of high unemployment may be periods of heightened uncertainty about the future, which may lead to reduced consumption and a higher incidence of consumption poverty. 10 TABLE 2 Estimated Effects of Unemployment and Inflation on Income Poverty, 1959-1992 Dependent Variable: Income Poverty Rate—Persons8 Explanatory Variables Constant Inflation (CPI-U) -0.26 -0.179 (0.192)c (0.146) -0.052 -0.015 (0.054) (0.054) Inflation (CPI-X1) Prime-age-male 0.4l7b unemployment rate (0.099) Demographic controls'1 Real hourly earnings 0.373b (0.104) Autocorrelation coefficient 0.50b (0.151) -0.81 l b (0.327) 0.335b (0.169) Adjusted R2 Number of observations 43.5% 57.7% 33 32 -1.62b (0.279) -0.036 (0.061) -1 .6 lb (0.275) -0.041 (0.056) 0.396b 0.390b (0.095) (0.095) yes yes n.a. n.a. 67.6% 67.9% 30 30 Dependent Variable: Consumption Poverty Rate—Persons* Explanatory Variables -0.389b (0.147) 0.207b (0.076) -0.262e (0.155) 0.180 (0.073) 0.453b unemployment rate (0.153) Demographic controls Real hourly earnings 0.317b (0.158) Constant Inflation (CPI-U) Inflation (CPI-X1) Prime-age-male Adjusted R2 Number of observations -1.349b -1.375b (0.414) (0.433) 0.204b (0.068) 0.21915 (0.088) 0.347b 0.322b (0.144) (0.150) yes yes -0.106b (0.485) 26.9% 35.7% 30 29 41.9% 36.4% 30 30 a. All data are first-differenced. b. Significandy different from zero at the 5 percent level or greater. c. Standard errors are in parentheses. d. Demographic controls include percent of persons over age 65, percent of white persons in population, and percent of families headed by a woman. e. Significandy different from zero at the 10 percent level or greater. SOURCE: Author’s calculations. Inflation may also be associated with height ened uncertainty and increased consumption poverty. Inflation can reduce permanent income (and hence consumption) by increasing the dis count rate applied to future income flows; this would also tend to increase the consumption http://fraser.stlouisfed.org/poverty rate. There are at least two other ways Federal Reserve Bank of St. Louis in which higher inflation might be associated with higher consumption poverty rates. First, inflation tends to benefit debtors at the expense of creditors, thus eroding asset values. Both liq uidity constraints and imperfect access to useful financial instruments may cause the net wealth of the consumption poor to be weakly hedged against inflation. Second, it is possible that households are slow to adjust their consump tion patterns to rapidly rising prices. This, too, might contribute to a higher rate of consump tion poverty. Rather than simply recomputing the regres sions reported in table 1 using the JS poverty rate in place of the conventional income poverty rate, all the data are first-differenced beforehand.19 This simple but effective method of detrending the variables is a special case of the flexible trend model employed in Mocan (1995). The top panel of table 2 presents the findings for the conventional poverty rate, and the bottom panel for the JS poverty rate.20 In the first column are the findings for the regression of the poverty rate on an intercept, the prime-age-male unemployment rate, and the growth of the CPI-U (with all variables firstdifferenced). For the conventional poverty rate, the findings remain qualitatively similar to those in the first column of table 1. The unemploy ment rate has a strong positive effect on pov erty, while the inflation rate has a negative, but statistically insignificant, effect.21 Both inflation and unemployment significantly increase the JS poverty rate. In contrast to the findings of Blank and Blinder (1986) and Blank (1993) that inflation’s effect is quite small relative to that of unemployment, the magnitude of the inflation effect on the JS poverty rate is nearly half that of the unemployment rate. The second column includes real wages, as suggested by Mocan (1995), who argues that if wage gains cause inflation, the effect of inflation on poverty may be biased downward when this variable is excluded. However, the findings indi cate that the estimated effect of inflation is robust with respect to the inclusion of real earn ings. The third column includes demographic variables (for age, race, and family type) that I 19 The model employs the same variables as Mocan (1995). It is noted below when the omission of variables from the models presented in table 1 affects the findings. ■ 20 The income-poverty-rate errors appear to follow an autoregres sive process of order one. I 21 Trend variables for the post-1962 and post-1983 periods were insignificant in the differenced specification and were dropped. 11 may have affected the overall incidence of poverty. The demographic variables are jointly significant. In both the conventional and JS poverty-rate specifications, the estimated coeffi cients are robust with respect to the inclusion of demographic variables, although the importance of inflation relative to unemployment in ex plaining the JS poverty rate grows even more pronounced.22 Due to the overstatement of inflation by the CPI-U and its possible contribu tion to overstating the conventional poverty rate, the alternative inflation rate based on the CPI-X1 was included, but the findings were not much affected. Overall, unemployment seems to have a strong positive influence on both poverty rates, while inflation is only influential for the JS poverty rate. The finding that unemployment increases the JS poverty rate suggests that either structural (long-run) unemployment is affecting the lifetime incomes of the poor, or that cyclical unemployment imposes permanent income losses. While Mocan (1995) presents evidence that the influence of unemployment on con ventional poverty rates is due to the adverse effects of long-run, not cyclical, unemployment, his findings are difficult to interpret, since the composition of the income poor is no doubt somewhat cyclical itself. In contrast, the esti mated effects of inflation on the two poverty rates are dramatically different. Inflation has a marginally negative effect on the conventional poverty rate, but a fairly large positive effect on consumption poverty. V. Conclusion This paper has reexamined the empirical rela tionships between inflation, unemployment, and poverty, using a methodology similar to that of previous work that apparently had shown the importance of unemployment and unimportance of inflation in influencing poverty rates. I have demonstrated that these previous findings are sensitive to seemingly reasonable alternative poverty measures. The findings presented here suggest that although unemployment’s effect on poverty rates is rela tively robust with respect to the poverty con cept, the effect of inflation on poverty may be more serious than previously thought. How should these new findings influence thought about the role of monetary policy? For those who subscribe to the view that the mone tary authority can lower or raise unemployment by enlarging or shrinking the money supply, the previous literature appeared to provide some evidence that expansionary monetary policy could make the average person better off by reducing unemployment, without the unpleasant side effect of making people worse off through inflation. The work presented here suggests that even if one accepts the existence of a trade-off between inflation and unemploy ment, one cannot be sanguine about the poten tial distributional costs of short-run stabilization policies, since the estimates are not robust with respect to alternative definitions of poverty. In further research, it would be interesting to decompose inflation into its anticipated and unanticipated components, and unemployment into its cyclical and long-run components. Unanticipated inflation might have the most adverse effects on consumption poverty if peo ple incorporate inflation expectations into their decisionmaking. It is also important to discover to what extent losses from transitory periods of high unemployment are made up in boom periods. Blank (1993) suggests that before the 1980s, low-income workers could make large real income gains during recoveries by increas ing their hours of work. In a consumption poverty framework, one would expect the cyclical effects of unemployment to be miti gated to the extent that these earnings gains are anticipated. However, there may be penalties for discontinuity in labor-force participation, implying that even cyclical unemployment could affect permanent income. It would be desirable to extend the data and analysis to examine the relationship between unemployment and inflation and the incidence of poverty within specific population sub groups. While the harmful impact of unemploy ment is still found to be larger than that of infla tion when consumption-based poverty measures are used for the entire population, it would be interesting to discover whether this qualitative finding is uniform across households, or whether a very strong effect of inflation on some, but not all, groups is driving the findings. Finally, the measurement of consumption poverty is a new and still controversial area. Based on their examinations of the CES sam ples (also used by Slesnick [1993D, Cutler and Katz (1991) conclude that “trends in the distri bution of consumption closely mirror those in ■ 22 The estimated coefficient of unemployment is highly sensitive to the inclusion of a government transfer variable in both the JS and con ventional poverty specifications, suggesting that innovations in govern ment policy and the prime-age-male unemployment rate are related. 12 the distribution of income” and that “while con sumption poverty rates are below income pov erty rates in every year, the time-series patterns for the two measures are quite similar.” Apply ing the Census equivalence scales and conven tional indexing to expenditure rather than income data, Slesnick finds the same pattern. It is his adjustment for the overstatement of infla tion and his and Jorgenson’s alternative equiva lence scales that generate the very different findings.23 Consequently, it is important to fur ther explore the extent to which the findings presented here are driven by specific assump tions employed in the construction of Slesnick’s consumption poverty rates. Nevertheless, the findings of this new research into the relationship between inflation, unemployment, and poverty have called the robustness of the earlier findings into question. More research is needed before we can confi dently say how macroeconomic developments affect poverty. References Blank, R.M. “Why Were Poverty Rates So High in the 1980s?” in Dimitri B. Papadimitrou and Edward N. Wolff, eds., Poverty a n d Prosper ity in the USA in the Late Twentieth Century. New York: St. Martin’s Press, Inc., 1993, pp. 21-55. ________ , and A.S. Blinder. “Macroeconomics, Income Distribution, and Poverty,” in Shel don H. Danziger and Daniel H. Weinberg, eds., Fighting Poverty: What Works an d What D oesn’t. Cambridge, Mass.: Harvard University Press, 1986, pp. 180-208. Blinder, A.S., and H.Y. Esaki. “Macroeconomic Activity and Income Distribution in the Post war United States,” Review o f Econom ics an d Statistics, vol. 6, no. 4 (November 1978), pp. 604-09. Cutler, D.M., and L.E Katz. “Macroeconomic Performance and the Disadvantaged,” Brookings Papers on Econom ic Activity, vol. 2 (1991), pp. 1-61. Danziger, S.H., and P. Gottschalk. “Do Rising Tides Lift All Boats? The Impact of Secular and Cyclical Changes on Poverty,” A m erican E conom ic Review, vol. 76, no. 2 (May 1986), pp. 405-10. Horrigan, M.W., and R.B. Mincy. “The Mini mum Wage and Earnings and Income In equality,” in Sheldon H. Danziger and Peter Gottschalk, eds., Uneven Tides: Rising Inequality in A m erica. New York: Russell Sage Foundation, 1993, pp. 251-75. Jorgenson, D.W., and D.T. Slesnick. “Aggregate Consumer Behavior and Household Equiva lence Scales,”Jo u rn al o f Business an d Eco nom ic Statistics, vol. 5 no. 2 (April 1987), pp. 219-32. _______ , and________. “Individual and Social Cost-of-Living Indexes,” in W. Erwin Diewert, ed., P rice Level M easurement. Amsterdam: North-Holland, 1990, pp. 155-234. I 23 Interestingly, the Jorgenson-Slesnick equivalence scales assume smaller efficiency gains in consumption as family size increases than do the Census scales. Thus, Slesnick argues that the Census meas ure actually understates poverty in the 1960s and earty 1970s. Minarik, J.J. “The Size Distribution of Income during Inflation,” Review o f Incom e an d Wealth, vol. 25, no. 4 (December 1979), pp. 377-92. 13 Mocan, H.N. “Income Inequality, Poverty, and Macroeconomic Conditions,” paper pre sented at the American Economic Association Meetings, Washington, D.C., January 7, 1995. Orshansky, M. “Counting the Poor: Another Look at the Poverty Profile,” Social Security Bulletin, vol. 51, no. 10 (October 1988), pp. 25-51. Powers, E.T. “Growth and Poverty Revisited,” Federal Reserve Bank of Cleveland, Eco nom ic Com m entary, April 15, 1995. Ruggles, P D raiving the Line: Alternative Poverty M easures an d Their Im plications fo r Public Policy. Washington, D.C.: The Urban Institute Press, 1990. Slesnick, D.T. “Gaining Ground: Poverty in the Postwar United States "Jou rn al o f P olitical Economy, vol. 101, no. 1 (February 1993), pp. 1-38. Yoshino, O. “Size Distribution of Workers’ Household Income and Macroeconomic Activities in Japan: 1963-88,” Review o f Incom e an d Wealth, vol. 39, no. 4 (Decem ber 1993), pp. 387-402. Scale Economies and Technological Change in Federal Reserve ACH Payment Processing by Paul W. Bauer and Diana Hancock Introduction Technological advances — accompanied by corresponding cultural changes and behavior adjustments — have had a tremendous influ ence on the array of payment instruments offered in the United States, on the diverse sys tems for processing them, and on their relative costs. Starting with the development of Mag netic Ink Character Recognition (MICR) in the 1950s, which facilitated the automation of check processing, the use of computers has trans formed virtually every aspect of banking and the payments system.1 For example, many new products, such as automated teller machines, point-of-sale terminals, touch-tone bill paying, and customer-initiated cash management serv ices, are now widely available. Advances in computer technology— speed, storage, commu nications, and encryption capabilities— have meant faster, more accurate, more secure, and less costly back-office processing. Since 1973, the use of electronic funds trans fers has been accelerated by development of the automated clearinghouse (ACH). The ACH system, a nationwide, value-dated electronic funds transfer system typically used for recur http://fraser.stlouisfed.org/ ring consumer and commercial payments, Federal Reserve Bank of St. Louis Paul w. Bauer is an economic ad visor in the Financial Services Research Group at the Federal Reserve Bank of Cleveland, and Diana Hancock is an economist in the Division of Reserve Bank Oper ations and Payment Systems at the Federal Reserve Board of Gover nors. For helpful comments, the authors thank Allen Berger, David Humphrey, Jeffrey Marquardt, and Florence Young. accommodates many types of transfers. The most common uses are to make utility, payroll, Social Security, tax, insurance premium, school tuition, mortgage, monthly investment, and divi dend payments, and to manage business’ cash concentration and disbursement activities.2 In the early 1980s, many observers argued that it would eventually become less expensive to transfer funds and settle most accounts elec tronically than to use traditional paper-based methods.3 Consistent with those expectations, the Federal Reserve’s direct and support costs for processing ACH payments today (approxi mately 1.4 cents per transaction) are less than 1 Payment data encoded at the bottom of checks have allowed high-speed check-sorting machines to process 80,000 to 100,000 checks per hour. ■ 2 The National Automated Clearing House Association (1995) esti mates that 42 percent of the private-sector workforce and 84 percent of government employees are paid using direct deposit. Also, more than 50 percent of Social Security recipients currently receive their benefits through direct deposit. ■ ■ 3 See, for example, Humphrey (1982,1984,1985). for paper checks (about 2.5 cents per check).4 Based on Federal Reserve data, the real unit cost (in 1994 dollars) of processing an ACH pay ment fe ll from 9.1 cents in 1979 to 1.4 cents in 1994. In contrast, the real unit cost of processing paper checks rose from 2.0 cents to 2.5 cents over the same period.5 Several hypotheses could account for the dramatic decline in both the absolute and the relative real costs of ACH processing. First, as the volume of ACH payments grew at double digit rates, per-item costs may have dropped because processing sites were able to achieve greater scale efficiency. By their basic nature, telecommunication systems, which consist of communication equipment and circuits, offer significant economies of scale over wide ranges of output.6 Such systems are one of the major inputs used in ACH payment processing. Early studies by Humphrey (1982, 1984, 1985), which used cross-sectional data from Federal Reserve ACH operations over the 1977-1982 period, verified that average ACH production costs fell as volume expanded. During that time, for each 1 percent rise in ACH processing volume, total production costs increased only 0.6 to 0.7 percent. Second, technological change may have made it cheaper to provide ACH services. With the same quantities of inputs, more funds trans fers could be processed. Software improve ments, for example, could have resulted in fewer computing resources being used to proc ess the same number of electronic payments. Third, some of the major inputs used for electronic payment processing, including com puters, experienced large quality-adjusted price declines during the 1980s. For the same cost, newer machines could process payments faster than their predecessors and could perform sophisticated tasks that were not previously feasible. Falling input prices would help to explain the absolute decline in real processing costs. At the same time, employee wages, paper costs, and other expenses associated with processing paper checks were generally rising.7 The change in relative input prices would help to explain the decline in the rela tive real unit costs of ACH processing. This study estimates the contribution of each of these factors— scale economies, technologi cal change, and falling input prices— to the absolute reduction in the real processing cost of an ACH transfer. We use Federal Reserve data over the 1979-1994 period and various specifications for ACH cost functions.8 Not sur prisingly, we find that all three factors played a significant role. The split between cost savings attributed to scale economies (through volume growth) versus technological change depends on the specification chosen for the cost func tion. While scale economies accounted for a decline in unit costs on the order of 20 to 40 percent, technological change explained more than 30 percent. Cost savings attributed to input price reductions generally accounted for less than 10 percent of the real per-unit decline in ACH payment processing costs. Our findings suggest that consolidating the Federal Reserve ACH processing sites will im prove scale efficiency, further reducing pro cessing costs. If recent experience is any guide, technological change will also present opportunities for further unit-cost declines. In addition, the marginal cost estimates presented in this study suggest that replacing paper checks with ACH transfers could enhance eco nomic efficiency. ■ 4 Direct and support costs cover all expenses specifically attribut able to providing Federal Reserve priced services, including labor, build ing, data processing, and data communication costs. They do not include allocations of overhead expenses, such as legal, accounting, and personal services, nor the Private-Sector Adjustment Factor (PSAF), which takes into account the taxes that would have been paid and the return on capital that would have been provided had the services been performed by a pri vate firm. Further, this definition of direct and support costs does not include the costs to payors and payees of processing payments. Thus, the Federal Reserve's costs are only a portion of the social costs of providing payment services. ■ 5 The real unit costs of processing ACH transfers and checks are calculated using the implicit GDP price deflator for 1979 and 1994. ■ 6 Scale economies were first studied in industries employing pipelines and boilers. There is a clear mathematical reason for this. Expanding the diameter of a pipe increases the amount of material required to manufacture it by only two-thirds as much as its capacity. (See, for example, Berndt [1991].) Similarly, in the context of communication sys tems, laying a fiber-optic line is not much costlier than laying a copper wire, but the former has many times the carrying capacity. ■ 7 Per-item wages have fallen over time because of capital improvements. ■ 8 The cost function is the minimum cost of producing any speci fied level of output given technological constraints and input prices. 16 I. What Is an ACH Transfer? Volume per quarter (millions of items) 1979 1981 1983 Volume per quarter (millions of items) 1985 1987 1989 1991 1993 The ACH system is a value-dated electronic funds transfer system. The principal participants in an ACH transaction are the payor, the payee, the payor’s bank, the payee’s bank, and the ACH operator.9 Either credit transfers or debit trans fers may be made using an ACH system. With credit transfers, such as direct payroll deposits, the payor’s bank typically initiates the transfer, and funds flow from the payor’s bank to the payee’s bank. With debit transfers, such as mortgage payments, the payee’s bank initiates the transfer and receives funds from the payor’s bank. ACH transactions offer several key advan tages over paper instruments. First, in most cases, payors know exactly when the funds will be removed from their accounts, and pay ees know exactly when the funds will be deposited to their accounts. Second, particu larly for consumer bill payments, ACH transac tions may be convenient because the payor does not have to remember to write and deliver a paper check, and the payee does not have to cash or deposit it. Third, the total costs to all parties are much lower for ACH transac tions than for paper checks.10 Finally, account ing efficiencies may exist for business payors and payees who have implemented electronic data interchange to facilitate communications with trading partners.11 1995 II. A Look at the Raw Data Before presenting statistical measures of scale economies and technological change, it is instructive to look at the raw data to determine how Federal Reserve ACH processing costs have varied over time and with different vol ume levels. Figure 1 presents unit costs (in 1994 dollars) over the 1979-1994 period, using ■ 9 We use the term "bank" to refer to all depository institutions. ■ 10 The full social cost of processing an ACH item is only about a third to a half as much as for a check (see Humphrey and Berger [1990] and Wells [1994]). 0 10 20 30 4 0 5 0 6 0 Volume per quarter (millions of items) SOURCE: Authors’ calculations. 7 0 8 0 9 0 ■ 11 See Knudson. Walton, and Young (1994) for a discussion of the potential benefits of financial electronic data interchange (a combina tion of electronic remittance data and electronic funds transfers) for busi ness payments. processing volumes as the measure of output.12 Despite improvements in the ACH service— including the introduction of encryption, in creased use of backup facilities, more deliveries per day, a wider variety of formats, provisions allowing more information to be supplied with the payment, and conversion to an all-electronic ACH environment— Federal Reserve per-unit costs have fallen steadily. Similar declines are observed at each processing site. For example, the Cleveland District’s unit-cost decline paral leled that of the System as a whole. This sug gests that technological change could have been the dominant factor driving down ACH process ing costs. However, output volume and input prices did not remain constant. Between 1979 and 1994, total ACH proc essing volume at the Federal Reserve grew at an average annual rate of more than 22 per cent (see figure 2), reaching 2.4 billion pay ments valued at $8.4 trillion by the end of 1994.13 If scale economies exist, then volume growth of this magnitude could account for a large share of the decline in unit costs. Figure 3 plots the unit cost per ACH transfer processed in the Cleveland Federal Reserve District against the number of quarterly trans fers processed at that site over the 1979-1994 period. Note that unit costs fell fairly steadily as volume increased. The experience at other Fed eral Reserve ACH processing sites was similar. Figure 3 suggests that scale economies (result ing from increased volume) were the dominant factor pushing down ACH processing costs. In general, however, output, technology, and input prices were all fluctuating (in some cases dra matically) over this period, necessitating a multi variate approach to data analysis to investigate changes in ACH costs. Both the formulation of public policies for electronic payments and the appropriate pricing framework for such pay ments hinge on an accurate understanding of the different sources of real unit-cost reductions. In general, the cost-function approach we employ in this paper is well suited to handling the contemporaneous effects of scale econo mies, as well as technological change and other factors. Unfortunately, the Federal Reserve s ACH data for each processing site show a strong correlation (greater than 99 percent) between the number of items processed (output) and a time trend (a technology index that is commonly used when a better measure is lacking). With such a high degree of correlation, it is difficult to disentangle the effects of technological change (the time trend) from those of scale economies (volume growth). From a technical standpoint, econometric models that include two highly correlated vari ables have upwardly biased standard errors, making it difficult to obtain precise estimates of the model’s parameters. Also, the cost-function coefficient estimates could be sensitive to small changes in the model’s specification. Since there is reason to believe that both scale economies and technological change are important factors in the real unit-cost decline for ACH processing, we choose to test for model robustness by trying alternative specifications for the cost function (for example, employing yearly indicators instead of a time trend to allow for technological change). We also use different sample periods within our pooled cross-section and time-series samples. III. Estimation To determine the effects of scale economies, technological change, and falling input prices on ACH processing costs, we estimate a cost function using quarterly cost data for Federal Reserve processing sites. This function maps the best (least-cost) method of processing each level of transfers when inputs, such as labor and computers, can be varied freely. In general, the least costly production method depends on the scale of operations. The cost function is a useful concept for our purposes, because many characteristics of technology can be derived from it, such as estimates of scale economies, marginal costs, and technological change (as will be explained more fully below). We employ the translog cost function be cause it provides a good local approximation of any arbitrary twice-differentiable cost function. Thus, the translog function can model many ■ 12 Throughout this paper, payments initiated and received at a processing site are counted once. Payments received and partially processed at one site and then transmitted and processed again at another are counted at both the sending and receiving sites. Theretore, processing volumes exceed the number of ACH payments made. ■ 13 A National Automated Clearing House Association press release dated March 27,1995 (“ACH Statistics Fact Sheet”) estimates that the total volume of payments handled by ACH processors (including the Federal Reserve) was 2.5 billion, valued at $10.1 trillion, in 1994. These statistics exclude estimated “on-us" items (wherein the payor and payee accounts are held at the same bank and consequently do not require exter nal processing). Although the growth rate and volume of ACH payments may seem impressive, these payments accounted for fewer than 4 percent of all noncash transactions processed domestically and only about 1 per cent of the dollars exchanged in 1994. 18 different possible relationships among the num ber of transfers processed (outputs), inputs, and environmental factors, depending on its para meter values. Our general translog cost func tion can be written as (1) InCtt = f io + j8yln> + 1/2 j8,.,(lny,,)2 K + X yMwki, k =i K K + 1/2 X k=i j X 8k] \nWkit\nWju =i K “I- X k =1 M Syk lny„ \viiVku 1994 + X ; = 1980 X \ mZ mn m= 1 (2) d2..— _ J? C _ dwkdw, dwjdwk 12 <bjYRj + X j= Zj Dj + Vu , 2 where is the number of ACH items proc essed at site i in period t, w„ is a vector of K input prices for site i in period t, Z„ is a vector of M environmental variables for site i in period t, Dt ( j =2,..., 12) is a set of site indicator variables (one for every processing site),14 YRj ( y =1980,...,1994) is a set of T-\ yearly indica tor variables (one for every year except the first), and v represents the error term.15 In some specifications, we use the time-trend term T= and its squared term, T 2, instead of the yearly indicator variables, YRj, to represent technological change. Depending on the model specification and the sample period selected, we set some of the coefficients of the translog cost function equal to zero. Several specifications of the cost func tion are estimated using ordinary least squares (OLS), and we denote these models as OLS Models 1 and 2. These elementary cost func tions include only an intercept, the log of the •number of items processed at each site, yearly indicators or a time-trend variable, and, in the case of OLS Model 2, some environmental vari ables. OLS Model 2 is similar to the cost func tion estimated by Humphrey (1982, 1984, 1985) for the ACH service. In our most sophisticated specification, we estimate the translog specification of the cost function jointly with the input share equations derived using Shephard’s Lemma.16 Estimation of both the cost function and the input share equations provides additional degrees of free dom and statistical precision. The system of cost and share equations is estimated using the iterative seemingly unrelated regression (ITSUR) technique.17 We denote these models as ITSUR Models 1 and 2. ITSUR Model 1 does not include site indicator variables (D,), in effect forcing the coefficients £,•(/= 2,..., 13) to equal zero. For both ITSUR models, we esti mate equation (1), along with the correspond ing equations for input shares, imposing the usual mathematical restrictions of symmetry and linear homogeneity in input prices. These restrictions, derived from economic theory, reduce the number of cost-function parameters that need to be estimated and thereby increase the number of degrees of freedom available. Symmetry restrictions follow from assuming that the cost function is twice differentiable in input prices, or This forces 8kl = 8,k for every k and j. Linear homogeneity in input prices means that only relative input prices matter. That is, propor tional changes in input prices affect only the level of cost, not the cost-minimizing set of inputs.18 Linear homogeneity restrictions result from defining the cost function as yielding the minimum cost of producing a given output level when faced with a particular set of input prices. In order to impose linear homogeneity, the following parameters related to the ln^*„’s are restricted such that Xk 7 * = (3) 1 and Xk 0 * = Xk 8 */ = 0. IV. Decomposition of Cost Savings over Time For a particular site, one could examine the ratio of unit costs in two periods. Although this ratio would show whether unit costs had risen or fallen, it would not indicate whether the shifi 14 The first processing site is the base against which the others are measured. Consequently, it does not have a site indicator variable. The choice of the base site does not affect our final results. ■ 15 The number of yearly indicator variables, YRt, depends on how many years of data are included in the sample. ■ ■ 16 See Diewert (1982) for a discussion of Shephard’s Lemma. 17 See Bauer and Hancock (1993) for a look at the various econo metric techniques that can be used to estimate a system of cost and share equations. ■ ■ 18 Mathematically, linear homogeneity can be expressed as \C(y, w) =C(y, X w), where X is greater than zero (x =2 if input prices double). 19 stemmed from scale economies, input price dif ferences, environmental differences, or techno logical change. To decompose the movements in unit costs attributable to various factors across time using cost-function (1), we can rewrite the ratio of a site’s current unit costs (with the period denoted by subscript S) to that of the first period (with the period denoted by subscript 0 ) as follows:19 C = In (Ks, We, ztf)exp(€g) Jfc C (y ,o , In w,o, Zto) exp (e,0) y-o Using the cost function defined in equation (1) and recalling that the log of a ratio is equal to the difference of the log of the numerator minus the log of the denominator, the percent age change in unit costs between periods, S and 0, or equation (4), can be rearranged into the following expression: (5) In (C, (C S S l'M )3y (lny,* - lny.o) + 1/2 P y y Q n t f s - In y*i0) - (InjVs - lny,(,) K I 7 * (In W ks- InWko) + k= i AC K + 1/2 I X 8k)(\nWkis\nWk,s k=\ j= i where the bracketed terms are defined as the technological change effects, scale effects (dif ferent processing volumes), input price effects (different input prices), interaction effects be tween processing volumes and input prices, environmental effects, and a random effect.20,21 Although these terms are in logarithmic differ ences, they can be roughly interpreted as the percentage difference in costs stemming from the various effects.22 Equation (5) provides a convenient framework for quantifying the source of cost savings over time. V. Data Construction We collected quarterly data from 1979 to 1994 on total costs, ACH processing volumes, input prices, and environmental variables for Federal Reserve ACH processing sites. During the 1979-1989 period, the number of these sites fell from 38 to 21. By an overwhelming mar gin, the largest volumes were handled by the 12 Reserve Banks and the Los Angeles branch of the San Francisco Fed. By 1993, only the 12 main Reserve Bank offices were still process ing ACH items. Consequently, we aggregated the data at the District level, with the exception of the Los Angeles facility, which we treated as a separate site. The New York Fed was omit ted from the estimations because most of the commercial ACH volume in its region was processed by the New York Automated Clear ing House. Our primary data source is quarterly cost accounting reports prepared by the Federal Reserve in its Planning and Control System (PACS). This information is supplemented by other cost and revenue data, results from occa sional Federal Reserve surveys, and price index figures from the Bureau of Economic Analysis (BEA) and the Bureau of Labor Statistics. Production costs for processed ACH transac tions are included in our calculations, but ■ 19 Any two periods could be chosen to compare unit costs, but comparing the first to the last is likely to be the most informative. - \nWk,o \nwkJo) ■ + ^dkdnyjnui.s - lny,olnMw)j k= ■ 21 The interaction effect is a collection of terms that cannot be classified cleanly into any of the other categories. Fortunately, the magni tude of this effect tends to be small. M + — m= 1 20 This decomposition uses the same methodology employed in Bauer (1993) to study differences in unit costs across Federal Reserve check-processing sites. Am (lnZm,s €,s — €,0 \nZmto) ■ 22 For the exact percentage difference, the antilog of each expres sion minus one should be used. We report the exact percentage differences of our results in table 5. 20 TABLE 1 Average Input Cost Shares, 1989-1994 (percent) Input Classification Cost Shares Labor Materials Communications Building 21.3 40.6 35.6 2.5 SOURCE: Authors’ calculations. imputed costs and certain overhead expenses, such as accounting costs and special District projects, are not. For the output measure, we use site-specific figures that focus on transac tions processed at a site, rather than the num ber of payments (see footnote 12). Labor, material, communication, and build ing costs are inputs for ACH processing. The shares of direct and support costs for each of these factors over the 1989-1994 period are reported in table l.23 Labor expenditures include salaries, retirement, and other benefits. The price of labor is total labor expenditures divided by the number of employee hours spent processing ACH transactions. While buildings’ share of costs is small, the interest expenses associated with the acquisi tion of fixed assets are not represented in the cost-accounting framework (these are included in the inputed costs [PSAF] rather than in direct and support costs). Cost accounting informa tion is supplemented by annual replacementcost indexes for each site, available from the R.S. Means Company.24 Square-foot replace ment costs, adjusted by the depreciation rate, are used to calculate maintenance and building prices for each site. Expenditures for materials are composed of outlays for office equipment and supplies, print ing and duplicating, and data processing. The service price for materials is constructed by sup plementing cost-accounting expenditure data with indexes for information and processing equipment.25 For computer hardware, an esti mate of the service value, or price, of machines is constructed using formulas that employ a per petual inventory model.26 For data system sup port services, which are primarily used for inhouse, product-specific software development, we construct a price by utilizing expenditures for labor and hours worked in that area of each Reserve Bank. For the service price of supplies (printing and duplicating, office supplies, and office equipment), we use the GDP implicit price deflator. We apply index number theory to construct a price index for materials that uses expenditures and prices for the components of materials— data processing, data systems sup port, and office supplies and equipment. Communications expenditures comprise the expenses associated with data and other com munications, shipping, and travel. The implicit price deflator for communications equipment purchases by nonresidential producers is used for data and other communications. The fixedweight aircraft price index for private purchases of producers’ durable equipment is employed for shipping and travel expenditures. Using index number theory, we calculate an overall price index for communications using the expenditure shares of two categories of com munications (communications and shipping) and their individual price indexes. Environmental variables that may affect ACH processing costs are the proportion of federal government items in the processing stream, the number of banks served by a processing site, and the proportion of banks receiving electronic payment information. On one hand, govern ment items may be less expensive to process because the Federal Reserve has more discre tion over file-processing times for these items than for commercial items. On the other hand, government items could be more expensive to process than commercial items because they are concentrated over short periods during the month and thus may drive processing capacity needs. The number of endpoints is the number of banks or processors to which ACH payments ■ 23 We focus on this period for several reasons. First, all of the data series are complete. Second, in the early period, full-cost pricing (required by the Monetary Control Act of 1980) was gradually introduced. Third, con solidation of processing sites could cloud the effects of scale economies in the early period. Consolidation effects are likely to be of minor significance, however, because of the low processing volumes and costs incurred at the additional sites. Finally, such dramatic technological changes occurred that a single cost function may be unable to fit the entire sample period ade quately. Consequently, by concentrating on the most recent data, we should get the best estimates of the current cost function for ACH processing. ■ 24 Data on replacement costs for buildings are taken from Means (1994). ■ 25 The BEAs implicit price deftator for information processing and related equipment is used for data processing and computer hardware. ■ 26 These formulas were derived by Hall and Jorgenson (1967). 21 t a b L E 2 Technological Change Indexes (1989 s 1.000) Year ITSUR Model 1 1989 1990 1991 1992 1993 1994 ITSUR Model 2 1.000 1.000 0.889 0.739 0.716 0.691 0.568 0.973 0.876 0.847 0.818 0.676 SOURCE: Authors’ calculations. D Q U R E 4 Technological Change Indexes Index, 1989 =1.00 SOURCE. Authors’ calculations. information is delivered. Nonelectronic deliver ies by computer tapes, diskettes, and paper methods increase transportation costs.27 In con trast, using electronic networks for deliveries may create greater scale efficiencies. VI. Empirical Results We estimated cost functions with and without the data for the early (1979-1988) period both to provide a historical perspective and to ease comparison with previous studies. The empiri cal results for the OLS cost-function models are reported in the appendix. Estimates from these models demonstrate that our qualitative findings are robust to changes in the assumptions employed in the estimation and in the sample period selected. In the body of the paper, we focus on the two ITSUR models estimated using http://fraser.stlouisfed.org/ data from 1989 to 1994. It is only during this Federal Reserve Bank of St. Louis period that data on the number of endpoints with electronic connections are available. Another reason we concentrate on the more recent period is that the methods used for ACH processing have changed dramatically over time. In the earlier period, ACH transaction data were delivered to the Federal Reserve Banks on computer tapes, and the Fed delivered data to receiving institutions on both computer tapes and paper listings. In the more recent period, however, ACH processing has essentially be come a computer network-based system. We are interested in whether different technologies for transmitting ACH transfers yield strikingly different estimates for scale economies and for technological change. Therefore, we estimate the cost function for the latest period possible — subject to the constraint of having sufficient degrees of freedom to estimate the model with statistical precision. ITSUR Models 1 and 2 estimate the cost func tion jointly with three of the four input share equations using the ITSUR technique. These models are preferred because they allow for a fuller complement of regressors and because including the cost-share equations increases sta tistical precision. ITSUR Model 2 differs from ITSUR Model 1 in that it includes processing-site indicator variables that allow for site-specific conditions not otherwise controlled for. Technological Change Table 2 presents estimates of technological change obtained from the two models above. The technological change index is set equal to one in 1989, with numbers below that indicat ing technological advance over the base year. For example, ITSUR Model l ’s 1994 index indi cates that unit costs are only 56.8 percent of costs in 1989, other things held constant. ITSUR Model 2 finds somewhat less technological change, with 1994 costs only 67.6 percent of those incurred in 1989. For ITSUR Model 1, this works out to a technological change estimate of more than 10 percent per year from 1989 to 1994. Inclusion of processing-site-specific inter cepts (ITSUR Model 2) lowers the estimate to just over 7.5 percent per year— still a rather hefty reduction. While the estimates of techno logical change differ, both models find the same pattern of unit-cost declines (see figure 4). ■ 27 All ACH transactions were delivered electronically as of July 1, 1993 for the commercial (non-federal government) sector and as of July 1, 1994 for the federal government sector. 22 TABLE 3 Cost Elasticity Estimates Federal Reserve ACH ITSUR Model 1 cessing Site 1994 0.764 0.776 0.761 0.771 0.766 0.763 0.760 0.768 0.765 0.761 0.772 1989 0.756 0.766 0.754 0.760 0.760 0.756 0.758 0.761 0.758 0.761 0.763 1 2 3 4 5 6 7 8 9 10 11 ITSUR Model 2 1989 1994 0.413 0.280 0.449 0.328 0.301 0.390 0.256 0.192 0.381 0.349 0.212 0.587 0.444 0.661 0.550 0.486 0.566 0.487 0.442 0.583 0.624 0.419 SOURCE: Authors’ calculations. F I GURE 5 Estimated Average Cost Functions Average cost per item (dollars) Volume per quarter (millions of items) SOURCE: Authors’ calculations. Scale Economies Cost elasticities measure the effect of a onepercentage-point increase in output on total cost. For example, a cost elasticity of 0.75 means that if output increases 1 percent, costs would rise only 0.75 percent. A cost elasticity of less than one indicates the existence of scale economies http://fraser.stlouisfed.org/ (that is, average cost falls as output increases). Federal Reserve Bank of St. Louis Alternatively, a cost elasticity greater than one indicates the existence of scale diseconomies (average cost rises as output increases). Although still finding significant scale econo mies during the 1989-1994 period, ITSUR Mod els 1 and 2 provide estimates of cost elasticities of widely different magnitudes. Table 3 presents cost elasticity estimates for each of the process ing sites that remained in operation for the entire sample period, using their mean processing vol ume levels for 1989 and 1994. ITSUR Model 2 provides greater estimates of scale economies for all sites than does ITSUR Model 1. To under stand why, consider figure 5, which plots the estimated average cost curves using both mod els. To the naked eye, these curves appear to be reasonably similar. In ITSUR Model 1, however, the coefficient of the squared term for the num ber of items processed is close to zero and is sta tistically insignificant. Based on this model, the cost elasticity is essentially constant at around 0.75 throughout the full range of observed out put, implying that scale economies are never exhausted. In contrast, for ITSUR Model 2, the squared term for the number of items processed is positive and statistically significant. This means that the cost elasticity varies along with the num ber of items processed. Consequently, ITSUR Model 2 suggests that scale economies will eventually be exhausted (that is, the average cost curve will eventually begin to rise). The volume level at which scale economies are exhausted is important, because it helps to determine whether consolidating the Federal Reserve processing sites could lower unit costs. Scale economies are exhausted when the cost elasticity equals one. By setting the cost elastic ity equal to one, we can solve for the implied number of items processed by a site operating at an efficient scale. Using this procedure and ITSUR Model 2, an estimate of about 800 mil lion items processed per quarter for a scaleefficient site is implied. This is more than five times the quarterly processing volume of the largest Federal Reserve site observed in our sample (144 million items per quarter). For both ITSUR models, estimates of the vol ume level at which scale economies are ex hausted need to be viewed with a fair degree of skepticism. Recall that the translog cost func tion is a good local approximation of the cost function and is therefore quite reliable in study ing output ranges actually observed in the data. While both models find significant scale econo mies in the current range of output, going be yond this range is highly speculative. 23 TABLE 4 Marginal Cost Estimates (dollars per item) Federal Reserve ACH Processing Site ITSUR Model 1 1994 1989 0.0204 0.0095 2 0.0215 3 4 0.0233 0.0182 0.0095 0.0087 0.0068 5 6 0.0269 0.0204 0.0135 0.0056 7 8 0.0321 0.0085 0.0215 0.0205 0.0262 0.0361 0.0069 0.0064 1 9 10 11 Volume-weighted System average 0.0234 ITSUR Model 2 1994 1989 0.0112 0.0079 0.0139 0.0078 0.0107 0.0105 0.0109 0.0054 0.0073 0.0054 0.0076 0.0049 0.0086 0.0042 0.0054 0.0040 0.0110 0.0103 0.0120 0.0101 0.0048 0.0062 0.0083 0.0106 0.0060 0.0076 0.0060 SOURCE: Authors’ calculations. The presence of scale economies implies that scale efficiency could be improved by con solidating the Federal Reserve ACH processing sites. Indeed, the Fed is currently consolidating its ACH operations at one computing site with backup facilities at another. Our empirical re sults suggest that these efforts will reduce aver age processing costs significantly. Comparing the average ACH processing cost at the current largest site with a forecasted average cost for a consolidated site handling all currently proc essed ACH items, the predicted average decline is 30 percent and 25 percent for ITSUR Models 1 and 2, respectively.28 Neither model predicts that scale economies would be exhausted with one processing site, but ITSUR Model 2 predicts that further scale efficiencies from additional volume growth could be quite small. Pricing The Monetary Control Act of 1980 directs the Federal Reserve to establish fees on the basis of all direct and indirect costs incurred in pro viding payment services, including “interest on items credited prior to actual collection, over head, and an allocation of imputed costs which takes into account the taxes that would http://fraser.stlouisfed.org/ have been paid and the return on capital that Federal Reserve Bank of St. Louis would have been provided had the services been provided by a private business firm.” Thus, the total revenues raised from providing payment services must match the total costs incurred in production. Generally, allocations of goods and services are most efficient when prices (the amount a consumer must pay to receive one unit of the good) are set equal to marginal costs (the addi tional cost of producing one more unit of out put).29 With scale economies of the magnitude we have found for ACH transactions, marginal cost pricing alone would not generate sufficient revenue to cover costs. The reason is that the presence of scale economies means unit costs fall as additional units are produced, and this can occur only if marginal costs are lower than average costs. The current Federal Reserve fee structure for the ACH service solves this prob lem by employing a multipart structure with both fixed and variable components.30 Ideally, to encourage greater use of electronic pay ments, the variable fee should be set equal to marginal costs and the fixed fees set to make up the shortfall. Our estimates of marginal costs, calculated using the two ITSUR models, are presented in table 4. Consistent with its finding of larger scale economies, ITSUR Model 2 generates lower marginal cost estimates than does ITSUR Model 1. Marginal costs for ITSUR Model 1 range from $0.0056 to $0.0135 per item in 1994, with a volume-weighted System average of $0.0083. ITSUR Model 2’s marginal costs are all estimated to be under $0.01 per item in 1994, with a volume-weighted System average of $0,006. Sources of Cost Savings In table 5, we use equation (5) to decompose unit-cost declines over the 1989-1994 period into technological change effects, scale econ omy effects, input price effects, environmental ■ 28 Note that with full consolidation, the number of items processed will equal the number of payments processed, approximately 600 million items per quarter. ■ 29 Mathematically, marginal cost (MC), the change in costs resulting from a unit increase in output, is defined as M C =dC /dy. ■ 30 See Baumol and Bradford (1970), Oi (1971), Roberts (1979), Humphrey (1984), Sheshinski (1986), Brown and Sibley (1986), Hirshleifer and Glazer (1992), and Tirole (1994) for discussions about efficient pricing methods when there are positive scale economies for an industry’s output level. 24 TABLE 5 Sources of Cost Savings ITSUR Model 1 Federal Reserve ACH Processing Site 1 2 3 4 5 6 7 8 9 10 11 Unit Cost51 1989 1994 0.027 0.028 0.012 0.012 0.011 0.009 0.018 0.007 0.011 0.009 0.008 0.010 0.014 0.031 0.024 0.035 0.027 0.042 0.028 0.027 0.034 0.047 Overall Technological Scale Economy Change Percentage Effects Effects Change Input Price Effects -43.2 -43.2 -43.2 -43.2 -43.2 -43.2 -43.2 -43.2 -43.2 -43.2 -43.2 -5 .8 -8 .3 -6 .0 -14.1 -9 .8 -9 .2 -6 .5 -8.1 -8 .8 -4 .0 -5 .3 -53.9 -56.5 -63.0 -62.9 -50.1 -72.7 -73.7 -68.1 -69.3 -71.0 -69.8 -16.9 - 1 6 .2 -20.2 -21.2 -18.4 -17.2 -23.0 -24.6 -19.6 -26.5 -20.4 Environmental Effects 0.1 -22.0 -5 .3 -5 .7 -5 .8 8.1 2.5 -28.3 -0 .8 1.3 -0 .5 Interaction Effects 0.5 -0 .3 1.0 0.3 -0 .4 0.2 -1 .0 -1 .4 0.5 0.5 -1.1 ITSUR Model 2 Federal Reserve ACH Processing Site 1 2 3 4 5 6 7 8 9 10 11 Unit Cost3 1989 1994 0.027 0.028 0.031 0.024 0.035 0.027 0.042 0.028 0.027 0.034 0.047 0.012 0.012 0.011 0.009 0.018 0.007 0.011 0.009 0.008 0.010 0.014 Overall Technological Scale Economy Percentage Change Effects Effects Change -53.9 -56.5 -63.0 -62.9 -50.1 -72.7 -73.7 -68.1 -69.3 -71.0 -69.8 -32.4 -32.4 -32.4 -32.4 -32.4 -32.4 -32.4 -32.4 -32.4 -32.4 -32.4 -37.5 -28.3 -46.8 -41.4 -34.2 -36.9 -40.5 -38.5 -41.3 -53.3 -32.2 Input Price Environmental Effects Effects -5 .9 -8 .3 -6 .0 -14.1 -9 .8 -9 .2 -6 .4 -8 .0 -8 .9 -4 .0 -5 .2 0.8 13.7 0.8 4.2 1.8 -1.1 1.8 18.9 2.3 0.1 2.5 Interaction Effects 0.6 -0 .4 1.1 0.3 -0 .4 0.3 -1 .0 -1 .5 0.6 0.5 -1.1 a. In dollars. SOURCE: Authors’ calculations. effects, and interaction effects. For each of the processing sites, unit costs fell precipitously. ITSUR Model 1 attributes the bulk of the decline, 43.2 percent, to technological change, versus only 32.4 percent for ITSUR Model 2. In contrast, ITSUR Model 2 finds larger cost savings due to scale economies than does ITSUR Model 1. Falling input prices— mainly for data com munications and data processing— generally account for less than 10 percent of the savings. As described in section V, we rely on the BEA’s price indexes to help construct our measure of materials, which includes information-processing and related equipment. The quality of such equipment changed rapidly during the 1980s and 1990s. Thus, to the extent that the price http://fraser.stlouisfed.org/ series for materials do not adequately control for Federal Reserve Bank of St. Louis the qualitative changes in these inputs, our decomposition of cost savings resulting from technological change may be overstated, while cost savings resulting from input price reduc tions may be understated. To some degree, the distinction is arbitrary. The decline in ACH costs may stem from technological change within ACH payments processing itself or from techno logical change in the computer industry that has lowered input prices. In either case, reduced costs from technological change are not misattributed to scale economies. Environmental and interaction-term effects tend to be relatively small, except for two sites, and these sites have by far the fewest number of endpoints. ITSUR Model 1 attributes their lower costs (other things held constant) only to 25 this factor. ITSUR Model 2, however, also allows for a different intercept term for these sites and finds smaller District indicator variable coeffi cients, suggesting that some other site-related factor is at work. with processing payments. Finally, in order to construct a pricing mechanism that encourages efficiency in the payments system and yet still recovers costs, the demand side of payment service markets— including cross-elasticities between payments instruments— needs to be more fully understood. VII. Conclusion We employ a cost-function model of ACH processing to derive estimates of both scale economies and technological change from 1979 to 1994. Substantial and statistically signif icant scale economies are found to exist at all Federal Reserve processing sites. For example, using cost system models, we estimate that for each 10 percent increase in ACH processing volume, total production costs rose by less than 8 percent, indicating that average costs fall as volume rises. Therefore, consolidating the System’s processing sites should reduce ACH processing costs substantially in the long run. In addition, given the potential scale economies for electronic payments processing and the low marginal costs, more attention is warranted for demand-side issues that would encourage payors to shift from paper checks to ACH transactions. More than 30 percent of the decline in real unit costs between 1989 and 1994 can be attrib uted to technological change, with an annual rate of change of at least 7.5 percent. Scale econ omies led to a further 20 to 40 percent reduc tion. Another significant contributing factor to the decline in unit costs was lower input prices (primarily for communications and computing technology), which translated into a cost savings of about 8 percent between 1989 and 1994. In the 1980s, some observers believed that scale economies would eventually push the interbank unit costs of processing ACH trans actions below those of processing paper checks.31 Our findings suggest that their expec tations were correct. In addition, technological change and lower prices for communications and computing technology have also played a major role. Clearly, more empirical research is needed on how new technologies affect the efficiency of the payments system. For example, scope economies between ACH payment processing and other payment processing, such as Fedwire and paper-based checks, could also be impor tant in determining the scale efficiency and optimal product mix for payment service providers. Such scope economies could enable many more suppliers to operate efficiently and to reduce the real resource costs associated Appendix: Sensitivity Analysis The strong correlation between the number of ACH items processed and the time-trend vari able makes it difficult to separate the effects of scale economies and technological change. Thus, we estimate equation (1) in several differ ent ways in order to determine the robustness of our qualitative findings of significant scale economies and technological change. First, we measure technological change either with a smooth time trend or with discrete yearly indi cator variables. Second, we estimate the model using data from the entire sample period, from each year separately, and from the 1989-1994 period only. OLS Model 1 is the most basic cost-function specification. Costs are regressed against the output measure (number of items processed) and against quarterly indicator variables. Models indicated by an a use yearly indicator variables to measure technological change, while models indicated by a b use a time trend and its squared term. OLS Model 2 is similar to OLS Model 1, except that control variables for the price of labor, the number of endpoints, and the percentage of government items are also in cluded. Model 2 is also similar to those esti mated by Humphrey (1982, 1984, 1985). A squared term for output is not included because the high correlation between the time-trend variable and output implies that the square of the latter could not be adequately handled in a single-equation setting. Below, we briefly sum marize the findings of these OLS estimations and compare them to the two ITSUR models presented in the body of the paper. ■ 31 For example, see Humphrey (1982,1984,1985) and Zimmerman (1981). 26 T A B L E A- 1 Technological Change Indexes (1989 = 1.000) OLS Model lb OLS Model 2a OLS Model 2b OLS Model la OLS Model lb OLS Model 2a OLS Model 2b ITSUR Model Year OLS Model la 1 ITSUR Model 2 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1.097 1.248 1.266 1.272 1.287 1.154 1.271 1.156 1.105 1.127 1.000 0.869 0.753 0.645 0.581 0.428 1.139 1.217 1.277 1.317 1.335 1.329 1.300 1.250 1.181 1.096 1.000 0.896 0.789 0.683 0.581 0.485 0.764 0.914 0.991 1.034 1.090 1.027 1.179 1.079 1.045 1.083 1.000 0.905 0.829 0.815 0.744 0.579 0.809 0.888 0.960 1.020 1.067 1.099 1.114 1.110 1.090 1.052 1.000 0.935 0.861 0.780 0.695 0.610 1.000 0.875 0.763 0.658 0.596 0.441 1.000 0.900 0.791 0.679 0.569 0.465 1.000 0.890 0.777 0.732 0.638 0.482 1.000 0.919 0.822 0.717 0.609 0.503 1.000 0.889 0.739 0.716 0.691 0.568 1.000 0.973 0.876 0.847 0.818 0.676 SOURCE: Authors’ calculations. F I G U R E A-1 Technological Change Indexes Index, 1989 =1.00 SOURCE: Authors’ calculations. Technological Change Estimates of technological change for all mod els are reported in table A-1 and plotted in figure A-1. To ease comparison, all of the technological change indexes are normalized http://fraser.stlouisfed.org/ to equal one in 1989. Whether yearly indicator Federal Reserve Bank of St. Louis variables or a time trend and its squared term are used to measure technological change, the technological indexes are similar in magnitude (see figure A-1 and compare Models a and b). This finding holds up only when data from the more recent period, 1989 to 1994, are em ployed. Basically, the time-trend approach reports a smoothed version of the yearly indi cator approach. The rise in the technological change indexes during the early 1980s could suggest technologi cal regress, defined as an upward shift in the cost function due to technological change. The technological change estimates from the early period (essentially, the start-up phase for ACH payments) are difficult to interpret, however, because it is plausible that the ACH cost func tion may have shifted substantially across time not only because of technological change, but also because of leaming-by-doing economies.32 ■ 32 Leaming-by-doing economies may have resulted from several factors. Workers performing repetitive tasks may have learned from cumu lative experience to perform these jobs more quickly and efficiently. Opera tions management at a processing site may have been able to call on its experience to modify job assignments, rearrange the layout of facilities, or devise ways to reduce paper or other material wastes In addition, software engineering may have improved computers' efficiency in processing batches of payments, so that the same amount of computer technology could process more ACH payments faster, or with greater security enhancement, and at lower cost. 27 T A B L E A- 2 Estimated Cost Elasticities [at sample means) Cost-Function Model Sample Period 1979-1994 OLS Model 1 Model a Model b 1989 1990 1991 1992 1993 1994 0.638a 0.7623 0.8523 0.897 0.968 0.8283 0.8493 0.6403 0.424a 0.523a 0.6713 0.6783 0.6483 0.748 o 00 o J9 8 9 -1 9 9 4 0.881a 0.8853 OLS Model 2 Model a Model b 0.851a 0.6343 0.6483 ITSUR Model 1 ITSUR Model 2 — — — — — — — — — — — — 0.7613 0.4483 a- Cost elasticity estimate is statistically different from one. SOURCE: Authors’ calculations. These cost-function shifts are difficult to model separately, particulady when a high cor relation exists between output and a time trend. In addition, we use year-specific indicator vari ables or a time trend and its squared term to derive estimates of technological change. Thus, it need not be the case that technological regress occurred: Other time-specific factors could also have increased ACH processing costs in the early years. Plausible candidates include the one-time transition costs to newer technologies, shifts to higher-quality (highercost) services with more bells and whistles, and various changes in cost-accounting procedures. Unfortunately, adequate control variables for such factors are unavailable, so we could not further decompose these time-specific effects. Estimates of technological change derived using models with more control variables tend to be larger, possibly because the models incor porate a greater number of environmental vari ables that control for site-specific characteristics. Scale Economies Estimates of cost elasticities at the sample means for the OLS models are reported in table A-2 for several different periods. Inclusion of additional site-specific regressors affects the estimates of scale economies, with OLS Model 2 yielding larger estimates (smaller cost elastici ties) than OLS Model 1. All of these cost elastic ities are statistically different from one at the 95 percent confidence level, confirming the pres ence of scale economies. Essentially, OLS Model 1 assigns more of the cost savings to technological change (and consequently, less to scale economies) than does OLS Model 2. Our estimates of cost elasticities are fairly close to the 0.70 to 0.75 figures reported in Humphrey (1982, 1984, 1985). Given the high degree of multicollinearity present in the data (particularly between output and a time trend), which over states standard errors and implies a bias toward rejecting the hypothesis of scale economies, a finding of statistically significant scale econo mies shows strong support for this hypothesis. As a further test of robustness, we also esti mate the two OLS models using quarterly crosssectional data for each year. Again, the scale economy estimates are larger when site-specific characteristics are included. Generally, the yearly cost elasticity estimates are statistically different from one at the 95 percent confidence level. Using OLS Model 1 in 1991 and 1992, 28 however, we do not find statistically significant scale economies. These estimates are bound to be less precise than those generated by our other models because they are based on very few observations. In summary, we subjected our cost-function model to a number of tests for robustness, pri marily by varying the sample period and the regressors. While the magnitude of some of the results varies significantly, our qualitative find ings across models are consistent. The sharp declines in unit cost appear to stem primarily from technological change and scale economies. Our finding of significant scale economies is ro bust to the model specification and selection of sample period. References Bauer, Paul W. “Efficiency and Technical Prog ress in Check Processing,” Federal Reserve Bank of Cleveland, Econom ic Review, vol. 29, no. 3 (Quarter 3 1993), pp. 24-38. ________ , and D. Hancock. “The Efficiency of Federal Reserve Check Processing Facilities,” Jo u rn al o f B anking an d Finance, vol. 17, nos. 2 and 3 (April 1993), pp. 287-311. Baumol, W. J., and D. Bradford. “Optimal Departures from Marginal Cost Pricing,” A m erican Econom ic Review, vol. 60, no. 3 (June 1970), pp. 265-83. Bemdt, E.R. The Practice o f Econom etrics: Classic a n d Contem porary. Reading, Mass.: Addison-Wesley Publishing Co., 1991. Brown, S., and D. Sibley. The Theory o f Utility Pricing. Cambridge: Cambridge University Press, 1986. Diewert, W.E. “Duality Approaches to Microeconomic Theory,” in K.J. Arrow and M.D. Intrilligator, eds., H an dbook o f M athem atical Econom ics, vol. 2. New York: North-Holland, 1982, pp. 535-99. Hall, R.E., and D.W. Jorgenson. “Tax Policy and Investment Behavior,” A m erican Eco nom ic Review, vol. 57, no. 3 (June 1967), pp. 391-414. Hirshleifer, J., and A. Glazer. Price Theory a n d Applications, 5th ed. Englewood Cliffs, N.J.: Prentice-Hall, 1992. Humphrey, D.B. “Costs, Scale Economies, Competition, and Product Mix in the U.S. Payments Mechanism,” Board of Governors of the Federal Reserve System, Staff Studies No. 115, April 1982. __________ . “The U.S. Payments System: Costs, Pricing, Competition, and Risk,” Mon ograph Series in F in an ce a n d Econom ics, New York University, Monograph 1984, nos. 1 and 2. _________ . “Resource Use in Federal Reserve Check and ACH Operations after Pricing,” Jo u rn a l o f B an k Research, vol. 16, no. 1 (Spring 1985), pp. 45-53- 29 _________ , and A.N. Berger. “Market Failure and Resource Use: Economic Incentives to Use Different Payment Instruments,” in D.B. Humphrey, ed., The U.S. Paym ent System: Efficiency, Risk, a n d the Role o f the F ederal Reserve. Boston: Kluwer Academic Publish ers, 1990, pp. 45-86. Knudson, S.E., J.K. Walton, and F. M. Young. “Business-to-Business Payments and the Role of Financial Electronic Data Inter change,” Board of Governors of the Federal Reserve System, F ederal Reserve Bulletin, vol. 80, no. 4 (April 1994), pp. 269-78. Means, R.S. M eans Square Foot Costs, 15th ed. Kingston, Mass.: R.S. Means Company, 1994. National Automated Clearing House Associa tion. “ACH Statistics Fact Sheet,” press release, March 27, 1995. Oi, W. Y. “A Disneyland Dilemma: Two-Part Tariffs for a Mickey Mouse Monopoly,” Q uarterly Jo u rn al o f Econom ics, vol. 85, no. 1 (February 1971), pp. 77-96. Roberts, K.W.S. “Welfare Considerations of Nonlinear Pricing,” E conom ic Jou rn al, vol. 89 (March 1979), pp. 66-83. Sheshinski, E. “Positive Second-Best Theory: A Brief Survey of the Theory of Ramsey Pricing,” in K. Arrow and M. Intrilligator, eds., H an dbook o f M athem atical Econom ics, vol. 3. New York: North-Holland, 1986, pp. 1251-80. Tirole, J. The Theory o f Industrial O rganiza tion, 7th ed. Cambridge, Mass.: MIT Press, 1994. Wells, K.E. “The Social Costs of Paper and Electronic Payments,” Virginia Polytechnic Institute, M.A. thesis, 1994. Zimmerman, G.C. “The Pricing of Federal Reserve Services under MCA,” Federal Reserve Bank of San Francisco, Econom ic Review, Winter 1981, pp. 22-40. 30 Economic Commentary Banking and the Flow of Funds: Are Banks Losing Market Share? by Katherine A. Samolyk September 1, 1994 Is Public Capital Productive? A Review of the Evidence by Kevin J. Lansing March 1, 1995 The Economics of Health Care Reform by Charles T. Carlstrom September 15, 1994 A Mexican Currency Board? by Owen F. Humpage March 15, 1995 Bank Receivership and Conservatorship by Walker F. Todd October 1, 1994 Are Wages Inflexible? by Ben Craig April 1, 1995 Specialization in Risk Management by Jerry L. Jordan October 15, 1994 Health Care Reform from a Generational Perspective by David Altig and Jagadeesh Gokhale April 15, 1995 Fear and Loathing in Executive Pay by Joseph G. Haubrich November 1, 1994 Understanding Differences in Regional Poverty Rates by Elizabeth T. Powers and Max Dupuy November 15, 1994 Can Foreign Exchange Intervention Signal Monetary Policy Changes? by William P. Osterberg May 1, 1995 How Much Is Daylight Credit Worth? by E.J. Stevens May 15, 1995 How Important Are U.S. Capital Flows into Mexico? by William P. Osterberg December 1, 1994 SAIF Policy Options by William P. Osterberg and James B. Thomson June 1995 (combines June 1 and June 15 issues) Year-End Report of the Fourth District Economists’ Roundtable by Michael F. Bryan and John B. Martin December 15, 1994 Monetary Policy and the Federal Funds Futures Market by John B. Carlson and Jean M. Mclntire July 1995 (combines July 1 and July 15 issues) Allocating Publicly Owned Assets: The Case of Personal Communications Services by Ian Gale January 1, 1995 Regulation and the Future of Banking by Jerry L. Jordan August 1, 1995 The Myth of the Overworked American by Kristin Roberts and Peter Rupert January 15, 1995 Another Look at Part-time Employment by Max Dupuy and Mark E. Schweitzer February 1, 1995 Monetary Policy: An Interpretation of 1994, a Challenge for 1995 by David Altig February 15, 1995 A Monetary Policy Paradox by Charles T. Carlstrom August 15, 1995 Derivative Mechanics: The CMO by Joseph G. Haubrich September 1, 1995 Should Social Security Be Privatized? by Jagadeesh Gokhale September 15, 1995 31 Economic Review ■ 1994 Quarter 3 I Restoring Generational Balance in U.S. Fiscal Policy: What Will It Take? by Alan J. Auerbach, Jagadeesh Gokhale, and Laurence J. Kotlikoff A Conference on Federal Credit Allocation by Joseph G. Haubrich and James B. Thomson Employment Creation and Destruction: An Analytical Review by Randall W. Eberts and Edward B. Montgomery Vagueness, Credibility, and Government Policy by Joseph G. Haubrich A Monte Carlo Examination of Bias Tests in Mortgage Lending by Paul W. Bauer and Brian A. Cromwell ■ Federal Funds Futures as an Indicator of Future Monetary Policy: A Primer by John B. Carlson, Jean M. Mclntire, and James B. Thomson 1994 Quarter 4 Tax Structure, Optimal Fiscal Policy, and the Business Cycle by Jang-Ting Guo and Kevin J. Lansing Cross-Lender Variation in Home Mortgage Lending by Robert B. Avery, Patricia E. Beeson, and Mark S. Sniderman The Efficiency and Welfare Effects of Tax Reform: Are Fewer Tax Brackets Better than More? by David Altig and Charles T. Carlstrom 1995 Quarter 1 I 1995 Quarter 2 An Introduction to Currency Boards by Owen F. Humpage and Jean M. Mclntire The Seasonality of Consumer Prices by Michael F. Bryan and Stephen G. Cecchetti 32 Third Quarter Working Papers Current Working Papers of the Cleveland Federal Reserve Bank are listed in each quar terly issue of the Economic Review. Copies of specific papers may be requested by completing and mailing the form below. Single copies of individual papers will be sent free of charge to those who request them. A mailing list service for personal subscribers, how ever, is not available. ■ 9507 Optimal Fiscal Policy When Public Capital Is Productive: A Business Cycle Perspective ■ 9509 Optimal Fiscal Policy, Public Capital, and the Productivity Slowdown by Kevin J. Lansing byStevenP.Cassouand Kevin J. Lansing ■ 9508 Marginal Tax Rates and Income Inequality: A QuantitativeTheoretic Analysis ■ 9510 Optimal Taxation of Capital Income in a Growth Model with Monopoly Profits by David Altig and Charles T.Carlstrom by Jang-Ting Guoand Kevin J. Lansing Institutional subscribers, such as libraries and other organi zations, will be placed on a mailing list upon request and will automatically receive Working Papers as they are published. Please complete and detach the form below and mail to: Research Department Federal Reserve Bank of Cleveland P.O. Box 6387 Cleveland, Ohio 44101 Check item(s) requested Send to: Please print City L J