The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Inflation Targeting William Poole This article was originally presented as a speech to Junior Achievement of Arkansas, Inc., Little Rock, Arkansas, February 16, 2006. Federal Reserve Bank of St. Louis Review, May/June 2006, 88(3), pp. 155-63. I am delighted to speak with Junior Achievement of Arkansas. I cannot report a story about how Junior Achievement (JA) got me off to a good start, but I do have a personal story—from my oldest son, Will. When I accepted this speaking invitation, I asked Will to reflect on his JA experience, and here is the paragraph he sent me. I was involved in Junior Achievement when I was in 8th grade. Most entrepreneurial-minded kids I knew gained their business experiences on paper routes, painting houses or the like. But I was drawn to JA’s concept of teaching the basics of business and figuring out how to mass-produce something. Little did I know where it would lead me. In my JA group, we assembled wooden coat pegs on boards and painted them up nicely. I quickly learned that building a single coat-rack widget is not so hard, but leading a handful of people to make 50, with quality, is much harder. And that getting all of them sold for a profit is even harder yet. I can’t say exactly which of the skills I learned at JA helped me end up running the Windows business at Microsoft. I was a big dreamer back then, but even I would not have dreamt that I would someday be leading a team of 3,000 professionals that create software that is used in 169 countries around the world and powers 200,000,000 new PCs sold every year. JA, thanks for the jump-start! Will is a senior vice president for Microsoft’s Windows client business. Needless to say, I am immensely proud of him. I don’t know the list, but will bet that numerous other JA alumni are in very responsible positions today. I find computers a bit mysterious, and I know that many think that monetary policy is even more mysterious. Federal Reserve officials used to delight in adding to the mystery, but today advances in macroeconomic theory have made clear the importance of central bank transparency to an effective monetary policy. Since coming to the St. Louis Fed in 1998, I have spoken often on the subject of the predictability of Federal Reserve policy, emphasizing that predictability enhances the effectiveness of policy.1 Predictability has many dimensions, but one is certainly that the market cannot predict what the Fed is going to do without a deep understanding of what the Fed is trying to do. The Fed has stated for many years that a key monetary policy objective is low and stable inflation. I believe that adding formality to that objective can clarify what the Fed does and why. That is my topic today. Before proceeding, I want to emphasize that the views I express here are mine and do not 1 See Poole (1999) for the first of a series of speeches on this topic. William Poole is the president of the Federal Reserve Bank of St. Louis. The author appreciates comments provided by colleagues at the Federal Reserve Bank of St. Louis. Daniel L. Thornton, vice president in the Research Division, provided special assistance. The views expressed do not necessarily reflect official positions of the Federal Reserve System. © 2006, The Federal Reserve Bank of St. Louis. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W M AY / J U N E 2006 155 Poole necessarily reflect official positions of the Federal Reserve System. I thank my colleagues at the Federal Reserve Bank of St. Louis for their comments; Daniel L. Thornton, vice president in the Research Division, provided special assistance. I take full responsibility for errors. THE FRAMEWORK The Federal Open Market Committee (FOMC) has the responsibility to determine monetary policy. The Committee implements policy by setting a target for the federal funds rate. Policy predictability does not mean that the public or the markets can successfully forecast the target federal funds rate next week, next month, or next year. The target rate is based on policymakers’ current information and best estimate of future economic events; the key observation is that incoming information may depart from the best estimate and indicate that the target funds rate needs to be changed to achieve policy objectives. What we must mean by perfectly predictable is that the public and the markets are not surprised by the Fed’s response to the latest economic information, understanding that the information itself is not predictable. Although new information creates a steady stream of mostly minor surprises, the FOMC ought to be clear about what it is trying to accomplish. At present, most members of the Committee would probably be pretty close together on how to state the inflation goal. A benefit of greater formality in defining the inflation goal is that individual FOMC members would have a clearer idea as to what the inflation objective is. To illustrate this point, I have often said that my preferred target rate of inflation is “zero, properly measured.” That is, allowing as best we can for measurement bias, which might be in the neighborhood of half a percent per year for broad measures of consumer prices, I favor literally zero inflation. Given measurement bias in price indices, I might state my goal as inflation between 0.5 and 1.5 percent as measured by the price index for personal consumption expenditures (the PCE price index). Others prefer a somewhat higher rate of inflation, perhaps in the range of 1 to 2 156 M AY / J U N E 2006 percent as measured by the PCE price index. Still others might favor a different target range, with a different midpoint and/or a wider or narrower range. If the FOMC decides to discuss inflation targeting, all dimensions of specifying a target will be considered carefully. Why does precision on a target range matter? Consider a situation in which the actual rate of inflation is 1.5 percent. Those favoring a target range of 1 to 2 percent would say that the policy stance is just right; inflation is in the exact center of the target range. I, given my preferred target range, would argue for a somewhat more restrictive stance, to move the inflation rate down toward the center of my preferred range. The difference between these two target ranges is small, and yet that difference might be enough to call for a somewhat different policy stance. Obviously, the Fed cannot simultaneously pursue two different inflation goals, and therefore there is every reason for the Committee to agree on a common objective. An agreed-upon common objective is much more important than the small difference between my own preferred objective and the range of objectives I believe are favored by others. If the FOMC were to decide on a common objective, then the Committee could communicate it to the general public. Discussion of the formal, numeric objective and what it means would help markets to better understand monetary policy and would make policy more predictable. However, many details matter and an inflation target will not be a source of increased clarity unless the details are specified appropriately. So, let’s talk about those important details. To simplify the language, I’ll refer to a publicly announced, specific numerical target range for inflation as a “formal” inflation target or objective. WHAT IS INFLATION? If the FOMC is going to adopt a formal inflation objective, we need to agree on what “inflation” is. However inflation is measured, it is important to distinguish between “high frequency” inflation, which central banks have little control over, and “low frequency” inflation, which central banks F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole can control. High-frequency inflation is the rate of change in the price level over relatively short time periods—months, quarters, or perhaps even a year. Low-frequency inflation is an economywide, systemic process that is affected by past, present, and expected future economic events. Central banks accept responsibility for lowfrequency inflation because such inflation depends critically on past and, especially, expected future monetary policy. When I advocate that the Fed establish a formal inflation objective, I am speaking of the low-frequency inflation rate. As a practical matter, low-frequency inflation can be thought of as the average inflation rate over a period of a few years. SETTING THE TARGET RATE OF INFLATION The Employment Act of 1946 sets objectives for monetary policy—indeed, objectives for all economic policy.2 The Act declares that it is the “responsibility of the Federal Government...to promote maximum employment, production, and purchasing power.” These objectives are reflected in the FOMC’s twin objectives of “price stability” and “maximum sustainable economic growth.” Although useful, these phrases are somewhat vague. For example, in the late 1970s and early 1980s, the Fed pursued the goal of price stability by reducing inflation from double-digit rates; from the mid-1980s into the early 1990s, the goal was to bring inflation down from the 4 percent neighborhood. Over the past decade or so, the goal has come to mean keeping the inflation rate low. But what inflation rate constitutes price stability? Rather than a numerical definition, former Chairman Greenspan preferred a conceptual definition, suggesting that “price stability is best thought of as an environment in which inflation is so low and stable over time that it does not materially enter into the decisions of households and firms.”3 But does Greenspan’s definition require zero inflation? 2 See Santoni (1986) for a discussion of the creation of the Act. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Because measuring the price level is a daunting task, zero true inflation and zero measured inflation may differ. Prices of individual goods and services change over time, but if some prices are falling and others are rising, then the average of all prices, or the price level, can remain constant. Nevertheless, defining a price index when prices are changing at different rates involves measurement issues that are complicated at both conceptual and practical levels. For a variety of technical reasons that I won’t discuss, the best we can do is to approximate the theoretical construct of the price level. Experts believe that price indices, such as the consumer price index (CPI) and the PCE price index, have an upward bias. That is, if the price level were truly unchanged, the price index would show a low rate of inflation. When asked during the July 1996 FOMC meeting what level of inflation does not cause distortions to economic decisionmaking, Chairman Greenspan responded, “zero, if inflation is properly measured.”4 Greenspan’s view that the theoretically correct definition of price stability is zero inflation stems from his belief that economic growth is maximized when the price level is unchanged on average over time.5 While I believe that there is a virtual consensus that the economy functions best when the theoretically correct measure of inflation is “low,” not everyone agrees with Greenspan that true price stability—a zero rate of inflation properly measured—is the best target for the Fed. For a variety of reasons, some economists believe that the economy functions best when inflation correctly measured is “low” but not zero. While the goal of price stability is specific in both the Federal Reserve Act and the Employment Act of 1946, some suggest that the FOMC lacks the authority to establish a numerical inflation objective. They claim that only Congress has this authority. That Congress has the power to estab3 Greenspan (2002, p. 6). 4 Transcript of the FOMC meeting held on July 2-3, 1996, p. 51. 5 For completeness, I note that Friedman (1969) argued that the optimal rate of inflation was negative. Specifically, he suggested that economic welfare was maximized when the nominal interest rate was zero. This requires that the inflation rate is equal to negative of the real interest rate. M AY / J U N E 2006 157 Poole lish the goals of economic policy is indisputable; however, it does not follow that the FOMC does not have the authority to adopt a formal inflation objective as part of implementing its broad congressional mandate. It is common practice for Congress to establish objectives and guidelines and leave it up to the agency responsible for meeting those objectives to fill in the details. The real question is this: Should the FOMC announce what its inflation objective is? Answering this question is simple in principle. If announcing a specific, numerical inflation objective enhances the efficacy of monetary policy, then the answer is yes. If doing so reduces the efficacy of monetary policy, the answer is no. I believe the answer is yes for a variety of reasons. THE CASE FOR AN INFLATION TARGET I have already pointed out that a formal inflation goal should improve the coherence of internal Fed deliberations by focusing attention on how to achieve an agreed goal rather than on the goal itself. Adopting and achieving a formal inflation objective should reduce risks for individuals and businesses when making long-term decisions. Because the benefits of price stability are indirect and diffuse, they are difficult to quantify. One area where the benefits of price stability are most apparent is the long-term bond market. It is not surprising that the 10-year Treasury bond yield has generally drifted down with actual and expected inflation since the late 1970s. The reduction in long-term bond yields reflects market participants’ expectations of lower inflation and their increased confidence about the long-term inflation rate. Moreover, the volatility of the market’s expected rate of inflation, measured by the spread between nominal and inflation-indexed 10-year Treasury bond yields, has trended down since the late 1990s, suggesting an increased confidence in the Fed’s resolve to keep inflation low. I anticipate that the adoption of a formal inflation objective would result in some, probably 158 M AY / J U N E 2006 modest, further reduction in the level and variability of nominal long-term bond yields. Adopting a formal inflation objective, and success in achieving that objective, will also enhance policymakers’ ability to pursue other policy objectives, such as conducting countercyclical monetary policy. I suspect that some of those who oppose a specific inflation objective are concerned that doing so will cause policymakers to become what Mervyn King, Governor of the Bank of England, has colorfully termed “inflation nutters.” King (1997) is referring to policymakers who aim to stabilize inflation, whatever the costs. I believe that just the opposite has happened. The debate is fundamentally about the relationship between the low-inflation objective and the high-employment objective. Even before British economist A.W. Phillips published research in 1958 that gave rise to what quickly came to be called the Phillips curve, many economists believed that there was a negative relationship between inflation and unemployment—i.e., lower inflation resulted in higher unemployment. Some preferred to think of causation as going the other way around—that higher unemployment resulted in lower inflation. The inflation-unemployment trade-off was thought to be permanent. Society could have a permanently lower average unemployment rate by accepting a higher average rate of inflation. In the late 1960s, Milton Friedman (1968) and Edmund Phelps (1967) challenged the idea of a permanent trade-off by making the theoretical argument that the Phillips curve must be vertical in the long run in a world where economic agents are rational. Subsequent evidence confirmed the Friedman-Phelps view, and few economists today believe that there is any long-run trade-off. A vertical long-run Phillips curve does not imply that one long-run inflation rate is as good as any other. Rather, the dynamics of the FriedmanPhelps theory imply that inflation would accelerate continuously were policymakers to pursue a policy of keeping the unemployment rate permanently below its natural, or equilibrium, rate. This equilibrium rate came to be called the NAIRU—the nonaccelerating inflation rate of unemployment. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole The Friedman-Phelps theory demonstrates why a policy of keeping the unemployment rate permanently below its natural rate is futile. It does not tell us the inflation rate that maximizes social welfare, which I will call the optimal inflation rate. Economic theory demonstrates why inflation is costly, and worldwide experience demonstrates that “high” inflation and “slow” economic growth appear to be inexorably linked. Everyone acknowledges that, beyond some rate, inflation reduces economic growth. The goals of price stability and maximum sustainable economic growth are not substitutes, as implied by the original Phillips curve, but complements. Monetary policymakers can make their greatest contribution to achieving maximum sustainable economic growth by achieving and maintaining low and stable inflation. That inflation and economic growth are complements does not imply that policymakers should not engage in countercyclical monetary policy when circumstances warrant. For example, with inflation well contained at the end of the long 1990s expansion, the FOMC began reducing its target for the federal funds rate in January 2001, somewhat in advance of the onset of the 2001 recession. The funds rate target was reduced still further in 2002 and 2003 as incoming data revealed that the economy was responding somewhat more slowly than expected and that actual and expected inflation remained well contained. The funds rate target was eventually reduced to 1 percent and remained there for slightly more than a year. Those who suggest that adopting a formal inflation objective will cause policymakers to become inflation nutters and, somehow, limit the Fed’s ability to pursue other policy objectives should examine actual experience. Not only did the Fed’s commitment to price stability not prevent it from engaging in countercyclical monetary policy—it facilitated it.6 Such an aggressive countercyclical monetary policy as pursued starting in early 2001 would have been unthinkable were it not for the fact that the credibility established over the years since Paul Volcker dramatically altered the course of monetary policy in October 1979.7 6 For evidence on how inflation interfered with countercyclical policy in the past, see Poole (2002). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W I believe that having a formal inflation objective will further enhance the Fed’s credibility and, consequently, its ability to engage in countercyclical monetary policy. The reason is simple. The more open and precise the Fed is about its long-run inflation objective, the more confident the public will be that the Fed will meet that objective. The objective, and the accompanying obligation to explain situations in which the objective is not achieved, should increase the Fed’s credibility. Because it will be much easier for the public to determine whether the FOMC is pursuing its inflation objective if that objective is known with precision, adopting a formal objective for inflation also will enhance the Fed’s accountability. Having a formal objective makes the Congress’s and the public’s job easier, thereby enhancing accountability. If the FOMC misses its inflation objective, it will have to say why the objective was missed. By the same token, the FOMC will have to explain why it failed to respond to a particular event when inflation appeared to be wellcontained within the objective. In essence, having a specific inflation objective will help the public better understand what I have elsewhere called “the Fed’s monetary policy rule.”8 SPECIFYING THE TARGET That there are differences of opinion about the optimal inflation rate is not a reason for having a fuzzy objective. If there are important differences of opinion within the FOMC on the appropriate target, which I doubt, the Committee ought to resolve those differences and not permit them to be a source of uncertainty. Because the target should apply to lowfrequency inflation, the target needs to be stated in terms of either a range or a point target with an understood range of fluctuation around the point target. The choice is more a matter of the most effective way of communicating the target and what it means than a matter of substance. 7 For those interested in understanding the issues that lead up to and succeeded this event, see Federal Reserve Bank of St. Louis (2005). 8 Poole (2006). M AY / J U N E 2006 159 Poole Figure 1A CPI and Core CPI 3-Year Moving Averages 12 10 8 6 4 2 0 1960 1965 1970 1975 1980 –2 M AY / J U N E 1990 1995 2000 2005 CPI for All Urban Consumers: All Items CPI for All Urban Consumers: All Items Less Food & Energy Difference A specific target range, such as 1 to 2 percent annual change in a particular price index, has the advantage of focusing attention on low-frequency inflation. Even here, there could be special circumstances, which the Fed should explain should they occur, that would justify departure from the target. The way the range is expressed interacts with the period over which inflation is averaged. A narrower range would be appropriate for a target expressed as a three-year average than for a year-over-year target. To understand what such a target means, suppose states were to abolish sales taxes and raise income taxes to offset the revenue loss. The effect of this change in tax structure would be to reduce measured prices. Such a tax change would be a one-time effect—the price level would change when the new tax law took effect but there would not be continuing pressure over time tending to lower prices. Suppose the one-shot price level change took measured inflation outside the target range. With a formal inflation target, the FOMC would have the responsibility of explaining why 160 1985 2006 a monetary policy response to this target miss would be unnecessary and perhaps harmful. A formal inflation target needs to refer to a particular price index. That there is no price index that adequately reflects the economy’s true rate of inflation is yet another reason given for not adopting a specific inflation objective. My own judgment is that the PCE price index measures consumer prices reasonably well and has some advantages, which can be explained, over the CPI. Moreover, the FOMC could reasonably maintain a rate of increase in this index in a range of, say 1 to 2 percent, on a two-year moving average basis under most circumstances. Over time, refinements in the price index or introduction of better indices may lead to substitution of another index for the PCE index or justify a change in the target range. The FOMC would then have to explain why it was adjusting the objective or index used to evaluate the objective. The formal target provides a valuable vehicle for explaining an important issue in the conduct of monetary policy. Experience with inflation targetF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole Figure 1B PCE and Core PCE 3-Year Moving Averages 10 8 6 4 2 0 1960 1965 1970 1975 1980 –2 ing in industrial economies suggests that issues of this sort have not been important. The markets are already well informed about such issues— and are becoming increasingly so. Conducting this conversation with the markets will improve the clarity of monetary policy and therefore its effectiveness. Over the past decade or so the Fed has gravitated to the position of placing primary emphasis on the core rate of inflation, as measured by the PCE price index excluding food and energy. The reason is not that food and energy are unimportant—these are obviously two very important categories of goods. Rather, experience indicates that food and energy prices are subject to large short-run disturbances that are beyond the ability of monetary policy to control without policy responses having adverse consequences for general economic stability. If we examine total and core price inflation over three years, say, most experience is that the averages are quite close. That is, food and energy prices display substantial shortrun variability that yields large changes in the F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W 1985 1990 1995 2000 2005 PCE: Chain-type Price Index PCE Less Food and Energy: Chain-type Price Index Difference short-run rate of inflation in overall price indices without affecting longer-run inflation. (See the charts in Figure 1, which track the CPI and PCE indices from 1960 through 2005.) HOW MUCH DIFFERENCE WOULD A FORMAL INFLATION TARGET MAKE? There is a large and growing literature comparing the performance of inflation-targeting countries with their non-inflation-targeting counterparts, especially the United States. This literature finds few statistically significant differences between countries that have established inflation targets and those that have not. This finding has led some analysts to argue, “if it isn’t broke, don’t fix it.” There are a number of reasons why such findings are not too surprising: The benefits from price stability are diffuse and difficult to measure; the industrialized economies are highly interconnected, so that some of the M AY / J U N E 2006 161 Poole benefits to countries that have inflation targets spill over to those that do not; the growth rate effect is small, so it will take a long time before one can distinguish a statistically significant growth-rate effect. Finally, many of the countries that adopted an inflation target had a history of inflation. Adopting a target was a manifestation of a societal commitment to bring down and keep down the rate of inflation. Given that the United States pursued a successful anti-inflation policy after 1979 without a formal target, and established a high degree of monetary credibility, there is no reason to expect to observe measurable effects from adopting a target now. Nevertheless, I cannot help reflecting on other cases in which low inflation prevailed but did not last. Consider U.S. policy errors of the type that occurred in the mid-to-late 1920s and in Japan in the late 1980s. In both of these instances, policymakers failed to respond to deflation. I believe that a formal inflation target would have focused attention on the policy mistake leading to deflation and would have increased public pressure on the central banks to respond more forcefully. Similarly, the Fed failed to tighten policy appropriately in the late 1960s as inflation began its ascent. In the early 1960s, as today, the Fed enjoyed a high degree of market confidence and inflation expectations were low. At that time, only a small minority of economists thought that monetary policy was “broken” in any important way, and thus the case for “fixing it” was minimal. Would a formal inflation target in 1960 have been an ironclad guarantee that the Great Inflation would never have happened? Surely not. Would it have helped? I believe that the answer is surely yes. CONCLUDING REMARKS Inflation targeting is an approach to monetary policy adopted by many countries, in most cases in the context of a societal effort to address undesirably high inflation. The United States, fortunately, is not dealing with an inflation problem at this time. The case for adopting an inflation target is that it should help to avoid inflation in the future and should increase the effectiveness of monetary policy in a low-inflation environment. 162 M AY / J U N E 2006 The increase in policy effectiveness should arise from two consequences of a formal system of inflation targeting. The first consequence is that the market will likely hold inflation expectations more firmly. The second, and probably more important, consequence is that the inflationtargeting framework provides a vehicle, or structure, within which the FOMC can better explain its monetary policy actions and the policy risks it must face. Inflation targeting should increase accountability not so much by keeping score of target hits and misses but rather by encouraging a much deeper understanding of how monetary policy decisions are made. That understanding depends on continuing FOMC communications with the markets and the public and FOMC willingness to listen as well as talk. REFERENCES Federal Reserve Bank of St. Louis. “Reflections on Monetary Policy 25 Years After October 1978: Proceedings of a Special Conference.” Federal Reserve Bank of St. Louis Review, March/April 2005, 87(2, Part 2). Friedman, Milton “The Role of Monetary Policy.” American Economic Review, March 1968, 58(1), pp. 1-17. Friedman, Milton. “The Optimum Quantity of Money,” in The Optimum Quantity of Money and Other Essays. Chicago: Aldine Publishing, 1969, pp. 1-50. Greenspan, Alan. Chairman’s Remarks. Federal Reserve Bank of St. Louis Review, July/August 2002, 84(4), pp. 5-6. King, Mervyn. “Changes in U.K. Monetary Policy: Rules and Discretion in Practice.” Journal of Monetary Economics, June 1997, 39(1), pp. 81-97. Phelps, Edmund S. “Phillips Curves, Expectations of Inflation and Optimal Employment over Time.” Economica, August 1967, 34(3), pp. 245-81. Phillips, A.W. “The Relation between Unemployment and the Rate of Change of Money Wage Rates in F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Poole the United Kingdom, 1861-1957.” Economica, November 1958, 25(100), pp. 283-99. Poole, William. “Synching, Not Sinking, the Markets.” Speech prepared for the meeting of the Philadelphia Council for Business Economics, Federal Reserve Bank of Philadelphia, Philadelphia, August 6, 1999; www.stlouisfed.org/news/speeches/ 1999/08_06_99.html. Poole, William. “Inflation, Recession and Fed Policy.” Speech prepared for the Midwest Economic Education Conference, St. Louis, April 11, 2002; www.stlouisfed.org/news/speeches/2002/ 04_11_02.html. Poole, William. “The Fed’s Monetary Policy Rule.” Federal Reserve Bank of St. Louis Review, January/ February 2006, 88(1), pp. 1-11; originally presented as a speech at the Cato Institute, Washington, DC, October 14, 2005. Santoni, G.J. “The Employment Act of 1946: Some History Notes.” Federal Reserve Bank of St. Louis Review, November 1986, 68(9), pp. 5-16. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W M AY / J U N E 2006 163 164 M AY / J U N E 2006 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W The Geography, Economics, and Politics of Lottery Adoption Cletus C. Coughlin, Thomas A. Garrett, and Rubén Hernández-Murillo Since New Hampshire introduced the first modern state-sponsored lottery in 1964, 41 other states plus the District of Columbia have adopted lotteries. Lottery ticket sales in the United States topped $48 billion in 2004, with state governments reaping nearly $14 billion in net lottery revenue. In this paper the authors attempt to answer the question of why some states have adopted lotteries and others have not. First, they establish a framework for analyzing the determination of public policies that highlights the roles of individual voters, interest groups, and politicians within a state as well as the influence of policies in neighboring states. The authors then introduce some general explanations for the adoption of a new tax that stress the role of economic development, fiscal health, election cycles, political parties, and geography. Next, because the lottery adoption decision is more than simply a tax decision, a number of factors specific to this decision are identified. State income, lottery adoption by neighboring states, the timing of elections, and the role of organized interest groups, especially the opposition of certain religious organizations, are significant factors explaining lottery adoption. Federal Reserve Bank of St. Louis Review, May/June 2006, 88(3), pp. 165-80. L otteries have had a turbulent history in the United States.1 In early America, lotteries were used by all 13 colonies to finance improvements in infrastructure, such as bridges and roads. Both during and after the Revolutionary War, lotteries were used to provide support for the military (e.g., the Continental Army), public projects, and the financing of private universities, such as Harvard. These early lotteries were closer to a raffle than to the modern concept of a lottery. Private lotteries began operating in the mid-1800s, with many of these lotteries operating through the mail system. As a result of corruption and a growing public distrust of lotteries, the federal government prohibited all interstate lottery commerce in the early 1890s. 1 See Clotfelter and Cook (1989 and 1990) for an extensive history of state lotteries. As a result of this federal prohibition and growing public distrust, the majority of states enacted explicit constitutional prohibitions against lotteries of any form. By 1894 no state allowed the operation of a lottery.2 Lotteries remained illegal in the United States for almost 70 years. In the early 1960s, however, New Hampshire had a lottery referendum that allowed the citizens of New Hampshire to vote for or against a state-sponsored lottery. Not only was New Hampshire the first state to propose the legalization of lottery gambling after 70 years of nationwide prohibition, it was the first modern attempt at state-run gambling. The voters of New Hampshire decided in favor of a lottery, with 76 percent of public votes in favor of adoption. In 1964, New Hampshire became the first state to 2 See Blanch (1949). Cletus C. Coughlin is deputy director of research, Thomas A. Garrett is a research officer, and Rubén Hernández-Murillo is a senior economist at the Federal Reserve Bank of St. Louis. Lesli Ott provided research assistance. © 2006, The Federal Reserve Bank of St. Louis. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W M AY / J U N E 2006 165 Coughlin, Garrett, Hernández-Murillo offer a lottery. Since that time, 41 additional states and the District of Columbia have adopted state-sponsored lotteries.3 North Carolina adopted a lottery in the summer of 2005, leaving only Alabama, Alaska, Arkansas, Hawaii, Mississippi, Nevada, Utah, and Wyoming without a lottery. Coinciding with the more frequent use of lotteries has been a rise in both lottery purchases and the importance of lottery revenue as a percentage of state revenue. Lottery ticket sales in 2004 topped $48 billion, or about 0.5 percent of total national income.4 Of this $48 billion in sales, states received nearly $14 billion in net lottery revenue (i.e., revenue available to state governments after the deduction of prizes, commissions, and administration costs).5 In terms of national per capita spending, lottery sales amounted to roughly $166 per person in 2004. Net lottery revenue as a share of total state government revenue rose from 0.35 percent in 1980 to 1.22 percent in 2002. The spread of state lotteries coincides with changing attitudes toward legalized gambling, growing state and local government expenditures, and growing public opposition to both new taxes and increased rates for existing taxes (Fisher, 1996). Arguably, lotteries are a more politically attractive means of generating additional revenue than increasing rates on existing tax bases. Although this premise may explain the initial interest in modern lotteries, it fails to adequately explain the uneven rate of lottery adoption over 3 Many states have entered into multistate lottery games, such as PowerBall. Multistate games pool ticket revenue from participating states to offer much larger jackpots (at more remote odds of winning) than single-state lottery games. Since 1964, states’ participation in multistate games has increased. One likely reason for this increased participation is to maintain players’ excitement about lottery games in the face of increased competition from casino gaming. See Hansen (2004) for a description of the various types of lottery games, including multistate lottery games. 4 Several states, such as Delaware and West Virginia, operate video lottery terminals at pari-mutuel racetracks. These venues are similar to casinos and generate hundreds of millions of dollars annually. Clearly, this form of lottery differs from the traditional scratch-off or numbers game lottery. Sales figures presented here include both traditional lottery sales and video lottery sales. 5 On average, states allocate 50 percent of sales to lottery prizes and 20 percent to administrative costs and retailer commissions. The remaining 30 percent is retained by the state. Hansen (2004) notes, however, that the shares for prizes and administrative costs vary by state. 166 M AY / J U N E 2006 the past 40 years (see Table 1). For example, between 1964 and 1975, 14 states adopted lotteries. No states adopted lotteries in the late 1970s, 18 states adopted lotteries in the 1980s, and 6 states adopted a lottery in the 1990s. In this paper, we focus on the question of why most states have adopted lotteries and why some states have yet to adopt a lottery. The growth in government and relaxed moral views of gambling may be a partial answer to this question, but these reasons are too broad, as they ignore the political and economic realities of public policy formulation. We review the literature on lottery adoption and, more importantly, public policy adoption in general to understand which factors drive policy formation. As the title of our paper suggests, lottery adoption is the result of geographic, economic, and political factors. THE POLITICAL ECONOMY OF PUBLIC POLICIES Public policy decisions result from the interaction of various factors. Before examining the factors that play a role in lottery adoption, we provide a framework for analyzing public policy decisions in general. Figure 1 highlights the primary actors and the legislative decision process in the democratic determination of a public policy.6 Similar to the demand and supply for a good, there also exists a demand side and a supply side for legislation, in this case the adoption of a state lottery. On the demand side, one starts with the opinions that individuals possess concerning the adoption of a lottery (see box A in Figure 1). The opinion of an individual is likely to be related to numerous considerations, such as income, education, age, potential impact of the legislation, and moral values. A common feature of any political decision in the United States is that interest groups are involved. Because of the intensity of opinions on an issue and the importance (economic and otherwise) of the issue, interest groups form and attempt 6 Our framework is based on ideas presented by Rodrik (1995) in the context of the political economy of trade policy. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Coughlin, Garrett, Hernández-Murillo Table 1 State Lottery Adoption State Arizona California Colorado Connecticut Delaware District of Columbia Florida Georgia Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland Massachusetts Michigan Minnesota Missouri Montana Nebraska New Hampshire New Jersey New Mexico New York North Carolina North Dakota Ohio Oklahoma Oregon Pennsylvania Rhode Island South Carolina South Dakota Tennessee Texas Vermont Virginia Washington West Virginia Wisconsin Start date Method of approval July 1, 1981 October 3, 1985 January 24, 1983 February 15, 1972 October 31, 1975 August 22, 1982 January 12, 1988 June 29, 1993 July 19, 1989 July 30, 1974 October 13, 1989 August 22, 1985 November 12, 1987 April 4, 1989 September 6, 1991 June 27, 1974 May 15, 1973 March 22, 1972 November 13, 1972 April 17, 1990 January 20, 1986 June 27, 1987 September 11, 1993 March 12, 1964 December 16, 1970 April 27, 1996 June 1, 1967 March 30, 2006 March 25, 2004 August 13, 1974 October 12, 2005 April 25, 1985 March 7, 1972 May 18, 1974 January 7, 2002 September 30, 1987 January 20, 2004 May 29, 1992 February 14, 1978 September 20, 1988 November 15, 1982 January 9, 1986 September 18, 1988 Initiative Initiative Initiative Legislation Legislation Initiative Referendum Referendum Referendum Legislation Referendum Legislation Referendum Referendum Referendum Referendum Referendum Legislation Referendum Referendum Referendum Referendum Referendum Legislation Referendum Legislation Referendum Legislation Referendum Legislation Referendum Initiative Legislation Referendum Referendum Referendum Referendum Referendum Referendum Referendum Legislation Referendum Referendum SOURCE: Hansen (2004); North Carolina and Oklahoma information from news reports. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W M AY / J U N E 2006 167 Coughlin, Garrett, Hernández-Murillo Figure 1 The Determination of a Public Policy: Lottery Adoption “Demand Side” Individual Opinions (A) combined with Interest Groups (B) Public Policy: Lottery Adoption/Rejection (F) Policies of Other States (E) Policymaker Opinions (C) combined with Institutional Structure of Government (D) “Supply Side” to influence the political decision (see box B in Figure 1). Through lobbying and contributions, interest groups attempt to affect the positions of representatives voting on the legislation.7 In addition, they attempt to increase popular support for their position as well. Thus, individual opinions and interest groups determine the demand side of the market for a public policy.8 On the supply side, one starts with the opinions of policymakers (see box C in Figure 1). These 7 McCormick and Tollison (1981) model an interest group economy using supply and demand analysis. Becker (1983) presents a theory of public policy formation that results from competition among special interest groups. 8 In Figure 1, we separate the demand side from the supply side for clarity in presentation. Because interest groups attempt to influence policymakers directly, we could have drawn a line from box B to box C. For illustrative purposes, one can think of such influence as playing itself out through the interaction of demand and supply. 168 M AY / J U N E 2006 policymakers include the legislators and those in the executive branch who can affect the legislation. Because the majority of these policymakers are elected representatives who, in many cases, wish to be re-elected, it is reasonable to anticipate that their positions will reflect to some degree the preferences of those who have elected them. As mentioned previously, interest groups also attempt to influence the positions of policymakers and are frequently involved in the drafting of legislation.9 The other consideration on the supply side is the institutional structure of government (see box D in Figure 1). Legislation is not simply proposed and then voted on, but rather it must work its way through a legislative process. As a piece of legislation is subjected to the scrutiny of legisla9 See Bonner (2005). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Coughlin, Garrett, Hernández-Murillo tive committees, the legislation may well be modified. The institutional structure of government might also affect the support for a piece of legislation by means of the trading of support that occurs between legislators. The extent to which a specific party is in control also might affect political support for the adoption of a lottery. Finally, the decisions made by policymakers and voters in a specific state may affect the decisions of other states (see box E in Figure 1).10 Geography as well as economics may come into play because the economic effects stemming from a specific state’s lottery may be more pronounced for nearby states. As a result, citizens and decisionmakers in nearby states may feel they must take a similar action (i.e., in this case, also adopt a lottery) as a form of self-defense. The prior adoption by another state may also provide information on the consequences of such legislation, which may influence the positions of individuals and policymakers. Ultimately, the various factors mentioned above interact to produce a decision. In the present case, the public policy decision is whether to adopt or reject a state lottery (see box F in Figure 1).11 As shown in Table 1, the precise method of approval of lotteries varies across states. Many states use a statewide referendum as part of the adoption process.12 A referendum is a popular vote on an issue already approved by a legislative body, with the final decision made by the electorate rather than by their representatives. Instead of a referendum, some states have adopted lotteries through the initiative process. The initiative process enables a specified number of voters to propose a law by petition. In the case of California, Proposition 37 was submitted to California’s voters, who approved the law. Finally, lottery adoption in several states, most recently North Carolina, simply required approval by each state’s 10 For example, Hernández-Murillo (2003) provides evidence of tax competition across states and Garrett, Wagner, and Wheelock (2005) present empirical evidence of cross-state dependence in banking deregulation. 11 Hersch and McDougall (1988) and Garrett (1999) explore legislative voting behavior on the issue of state lottery adoption. 12 Hansen (2004) notes that many states required a referendum or an initiative to remove a state constitutional ban on lotteries. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W legislature and governor without a direct citizen vote.13 THE ADOPTION OF NEW TAXES Without question, the issues surrounding the adoption of a tax are similar to those for the adoption of a lottery; however, the adoption of a lottery entails more than (and is quite different from) simply authorizing a new tax. Lottery adoption involves legalizing a previously illegal activity, from which a state will generate revenue. Furthermore, consumer participation in the lottery is strictly voluntary; it is possible, then, that those who are opposed to the lottery might not oppose the legislation because they can decline to play the lottery, whereas it is more complicated and potentially illegal to decline to pay a tax.14 Also, one’s position on lottery adoption as a means to raise revenues might be overwhelmed by other considerations. Therefore, our discussion of lottery adoption must go beyond an explanation of adopting a new tax: In fact, our discussion integrates all these considerations and others after examining some literature on the adoption of new taxes. Hansen (1983) develops a theory of taxation highlighting the role of political incentives, many of which apply to the lottery adoption process. According to Hansen, politicians make decisions with an eye toward retaining their positions. Economic considerations come into play in the adoption of new taxes by affecting the political incentives. For example, the existence of an economic crisis may reduce the political risks of approving new taxes, whereas new taxes are unlikely to gain approval if a state has a budgetary surplus on the horizon. In the absence of crises, separating taxpayers from their incomes/wealth is an unwise electoral strategy. Even with a crisis, however, Hansen stresses the importance of politi13 The initiative is similar to a referendum; however, policymakers have a more limited role in the approval of lottery adoption if the initiative process is used, compared with either a referendum or standard legislative process. 14 Arguably, paying sales and income taxes is voluntary if one chooses not to purchase consumer goods or work. M AY / J U N E 2006 169 Coughlin, Garrett, Hernández-Murillo cal parties’ control of government for implementing tax policy. A unified government is crucial for providing the political opportunity for adopting a new tax. Capitalizing on political opportunity is an issue Berry and Berry (1992) take up. They identify the following categories of explanations for the adoption of new taxes, or what is often termed as a tax innovation: (i) economic development, (ii) state fiscal health, (iii) election cycles, (iv) political party control, and (v) regional diffusion. The economic development explanation suggests that a state’s level of economic development affects the likelihood of adopting a tax. More-developed states are likely to have a combination of tax capacity and demand for public services that lead to tax adoption.15 The fiscal health explanation suggests that the existence of a fiscal crisis, such as a large budget deficit, increases the probability of approving a new tax. The crisis reduces the political risks for politicians of a tax innovation. Similar reasoning is used in the election cycle explanation. Tax increases are unpopular; therefore, elected officials do not innovate in election years. The party control explanation has two propositions. First, if the party in control is a liberal party, the adoption of a new tax is more likely. Second, a state in which the same party controls the governorship and the legislative bodies (i.e., a unified government) is more likely to adopt a tax than a state with a divided government. The fifth explanation, the regional diffusion explanation, suggests that states emulate the tax policies adopted by others. Political scientists have stressed that prior adoptions provide information and make the tax increase easier to sell to constituents. In addition, in the context of lotteries, adoption by a state puts competitive pressures on nearby states because some of its lottery tax revenues are due to attracting players from nearby states. 15 As described in Filer, Moak, and Uze (1988), the Advisory Commission on Intergovernmental Relations uses a “representative tax system” to calculate tax capacity. Tax capacity is the revenue that each state would raise if it applied a uniform set of rates to a common set of tax bases. The uniform set of rates is the average rates across states for each of 26 taxes. Using the same rates for every state causes potential tax revenue for states to vary only because of differences in underlying tax bases. A state’s “tax capacity index” is its per capita tax capacity divided by the average for all states. 170 M AY / J U N E 2006 A sixth category for explaining the adoption of a lottery not discussed by Berry and Berry (1992) considers the alternatives and constraints facing policymakers and the factors that apply specifically to the approval of a lottery. We use the term “situational-specific determinants” to describe this category. When faced with pressures to increase revenues, policymakers examine a range of possibilities that include increasing the rates of existing taxes, expanding what is taxable, as well as adopting and implementing new taxes.16 The ability of policymakers to increase revenues may be limited by prior political decisions and by a state’s economic circumstances that are beyond economic development and fiscal health. For example, a state that has no sales tax is likely to face different political constraints than a state with a sales tax. Increasing a sales tax rate from 0 to 1 percent, which requires adopting and implementing a sales tax, is likely much different from increasing a sales tax rate from 4 to 5 percent. FRAMEWORKS FOR ANALYZING THE LOTTERY ADOPTION DECISION Various conceptual frameworks have been used to model state lottery adoption, but they all rely on rational behavior by legislators. What differentiates the frameworks is the objective function of the legislator. The most frequently used framework is the legislator-support maximization approach (see Filer, Moak, and Uze, 1988). The position on lottery adoption that a given legislator takes reflects an attempt to maximize re-election prospects. Legislators recognize that increased state spending can increase their political support by increasing the well-being of their constituents. Note that it is through political support that the demand side of the determination of a public policy is incorporated. Spending cannot be raised, however, without some loss of support due to the increased tax burden placed on their constituents. This trade-off guides how the legislator votes on specific issues. For a legislator to vote in favor of 16 Of course, cutting spending is also an option to deal with an imbalance between spending and revenues. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Coughlin, Garrett, Hernández-Murillo a specific proposal, the increase in support associated with the spending must exceed the decrease in support associated with the taxes. A second framework, Martin and Yandle’s (1990) duopoly transfer mechanism approach, views the state government as a rent-seeker and a redistributive agent. Lotteries compete with both legal and illegal gambling operations. A state-run lottery provides a mechanism that allows the state to generate some revenues that they miss by not being able to tax illegal operations. A closely related issue is why lotteries are organized as a state enterprise as opposed to allowing private firms to freely enter and provide lottery services in a competitive environment. One answer is that the revenues for the state from a state-run lottery are likely to exceed the revenues that the state would generate from allowing private firms to enter the lottery market and then taxing the profits of these firms. In Martin and Yandle’s (1990) approach, the state achieves equilibrium with the illegal operators. Moreover, lotteries provide a way for higher-income voters to redistribute the tax burden associated with state spending from themselves to lower-income groups.17 A third framework, in Erekson et al. (1999), assumes the legislator maximizes utility subject to a constraint. The legislator reflects the median voter because of the decisive role of this voter in producing a majority. The legislator receives utility from improving the state’s fiscal well-being, but the legislator is constrained, similar to the constraint in the legislator-support maximization approach, by his re-election desires that hinge on the satisfaction of his constituents. The empirical implementation of these frameworks has proceeded in two ways. Filer, Moak, and Uze (1988), Martin and Yandle (1990), and Davis, Filer, and Moak (1992) address the question of whether a state has a lottery as of a specific year. Filer, Moak, and Uze (1988) and Davis, Filer, and Moak (1992) estimate binary choice probit models, while Martin and Yandle (1990) use ordinary least squares. These studies identify and then examine statistically a number of variables that are related to whether a state has a lottery as of a 17 Note that, despite the focus on state government, there is still a demand role played by the state’s citizens. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W specific year. The other approach, using hazard or duration models, provides evidence on which variables increase or decrease the likelihood that a state adopts a lottery. Using this approach, the variable to be explained is termed the hazard rate, which is the probability that a state without a lottery will in fact adopt a lottery during a specific time period, generally a calendar year.18 WHY SOME STATES HAVE LOTTERIES AND OTHERS DO NOT Using the tax innovation explanations put forth by Berry and Berry (1992) that were discussed earlier, this section sheds light on the question of why some states have lotteries and others do not. Economic Development With respect to the level of economic development, higher levels of per capita state income are associated with (i) an increased probability that a state has a lottery as of a specific date and (ii) a shorter time until a state adopts a lottery.19 A common argument rationalizing this result is based on the finding that lotteries tend to be a regressive form of taxation. In other words, evidence suggests that those with low incomes bear a relatively higher lottery tax burden than those with high incomes.20 In their report to the National Gambling Impact Study Commission, Clotfelter et al. (1999) provide evidence that low-income groups spend a larger share of their incomes on the lottery and that they also spent more in absolute terms.21 For example, those with an annual house18 Berry and Berry (1990), Alm, McKee, and Skidmore (1993), Caudill et al. (1995), Mixon et al. (1997), Erekson et al. (1999), and Glickman and Painter (2004) estimate hazard functions in their lottery adoption studies. 19 See Davis et al. (1992), Martin and Yandle (1990), Berry and Berry (1990), Caudill et al. (1995), Mixon et al. (1997), Erekson et al. (1999), and Glickman and Painter (2004). 20 There is some evidence that high jackpot lottery games, such as PowerBall, may be less regressive than lower jackpot games. See Oster (2004). 21 This latter finding suggests that lottery tickets are inferior goods (e.g., the income elasticity of demand for lottery tickets is negative). Although most studies have found lotteries to be regressive, most have not found lotteries to be inferior goods. See Clotfelter and Cook (1989 and 1990) and Fink, Marco, and Rork (2004) for a survey of the literature. M AY / J U N E 2006 171 Coughlin, Garrett, Hernández-Murillo hold income of less than $10,000 spent $597 on lotteries on a per capita basis and those with a household income of between $10,000 and $24,999 spent $569. This spending was substantially more than spending by those with a household income of between $25,000 and $49,999 ($382), between $50,000 and $99,999 ($225), and over $100,000 ($196). In light of the regressive nature of lotteries, it has been argued that a legislator with a lowincome constituency is more likely to oppose raising funds by means of a lottery than a legislator with a high-income constituency.22 Martin and Yandle (1990) stress this redistributive feature by arguing that lotteries are a mechanism for higherincome voters to redistribute tax burdens in their favor. Thus, states with higher per capita incomes are more likely to have lotteries.23 In addition to the redistribution argument, Berry and Berry (1990) and others stress that higher per capita income is associated with more revenue potential from a lottery. However, if higher-income households in fact spend less on lotteries than lowerincome households, then the revenue potential from a lottery may decline as per capita income rises. The connection between revenue potential and lottery adoption has been explored by Filer, Moak, and Uze (1988). They argue that states with a larger urban population are more likely to have lotteries than more rural states because of relatively lower administrative costs.24 More densely populated states will tend to have more potential purchasers per lottery outlet and generate a relatively greater value of revenue per dollar of administrative costs. Filer, Moak, and Uze find that the percentage of a state’s population that is urban is related positively to lottery adoption, while Alm, McKee, and Skidmore (1993), Caudill et al. (1995), and Glickman and Painter (2004) find that state population density is related positively to lottery adoption. However, Caudill et al. (1995) and Mixon et al. (1997) fail to find a statistically significant relationship between a measure of predicted lottery profits and lottery adoption. The empirical evidence examining the connection between total population and lottery adoption is also mixed. Alm, McKee, and Skidmore (1993) find a positive and statistically significant relationship between state population and lottery adoption, but Filer, Moak, and Uze (1988) and Glickman and Painter (2004) do no find a statistically significant relationship. Another measure of economic development that has been examined for its statistical relationship to lottery adoption is per-pupil state education spending. Lottery adoption often occurs as part of a promise to earmark lottery proceeds to finance spending on education. Thus, states with lagging education spending are more likely to support such earmarking. Second, the current lack of education spending portends a bleak economic future that motivates a state to take action to alter the future. Erekson et al. (1999) find a statistically significant, negative relationship between per-pupil education spending and the decision to adopt a lottery.25 22 This argument assumes that legislators with low-income constituencies take such a position for paternalistic reasons (rather than from a desire to represent the views of their constituents). Legislators with high-income constituencies, on the other hand, take a position in line with the views of their constituents. A related argument is that low-income groups with poor economic prospects may place a relatively higher discount on lottery losses than high-income groups. Lotteries offer a small prospect for a large gain for those who play. Representatives of low-income constituents could vote to restrict lotteries to inhibit these constituents from gambling away their minimal resources. The fiscal health explanation suggests that the existence of a fiscal crisis, such as a large budget deficit, increases the probability of approving a new tax or, similarly, a state lottery. Numerous variables have been used to measure a state’s fiscal health. A measure used by two early studies was 23 Filer, Moak, and Uze (1988) use a measure of the percentage of poor within a state and find this measure to be a negative and statistically significant determinant of lottery adoption. 24 DeBoer (1985) examines the economies of scale in state lottery production. 172 M AY / J U N E 2006 Fiscal Health 25 Although net lottery revenue is earmarked for public education in many states, there is little evidence that the earmarking of lottery revenue has increased education expenditures. The reason for this is that state legislators divert funds away from education and simply replace these diverted funds with net lottery revenues, thus leaving total education expenditures unchanged. See Spindler (1995) and Garrett (2001). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Coughlin, Garrett, Hernández-Murillo the Advisory Commission on Intergovernmental Relations’ tax effort. This measure is a ratio of a state’s tax collections relative to its capacity of tax revenue. Larger values of this measure are thought to indicate that a state was making an increased effort to generate tax revenues. As this measure becomes larger, the odds increase that a state will seek additional revenue sources. Both Filer, Moak, and Uze (1988) and Davis, Filer, and Moak (1992) find that the higher a state’s tax effort, the larger the probability the state has a lottery. Rather than use tax effort, some studies have simply used per capita state tax revenues. Martin and Yandle (1990) argue that higher per capita state taxes are an indicator of tax pressures, and, thus, higher levels provide an increased incentive to find a way to relieve the pressure. Martin and Yandle (1990) argue that one way to simultaneously relieve the pressure and increase tax revenues is to shift the relative tax burden from higher- to lower-income taxpayers by means of a lottery. They found that higher per capita state taxes were associated with lottery adoption. On the other hand, Caudill et al. (1995) and Mixon et al. (1997) do not find a statistically significant relationship between per capita state taxes and the probability that a state adopts a lottery in a given time period. Another commonly used measure of fiscal health is per capita state debt. Higher levels of debt raise increased doubts about a state’s fiscal health. As put forth in Martin and Yandle (1990), higher debt will result in a larger demand for shifting taxes to lower-income groups; such pressures make lotteries more likely. Martin and Yandle (1990) find a positive and statistically significant relationship between per capita state debt and lottery adoption. However, numerous other studies, such as Caudill et al. (1995), Mixon et al. (1997), and Glickman and Painter (2004), do not find a statistically significant relationship between per capita state debt and the probability that a state will adopt a lottery within a given period. On the other hand, when Alm, McKee, and Skidmore (1993) separate overall debt into short-term and long-term debt, they find a statistically significant, positive relationship between short-term state debt and lottery adoption, but no such statistically F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W significant relationship between long-term state debt and lottery adoption. Rather than looking at overall debt, Berry and Berry (1990) and Erekson et al. (1999) examine the difference between state revenue and spending relative to spending as a measure of fiscal health. This measure is a proxy for the budget deficits faced by states. The more negative this measure, the worse a state’s fiscal health. It is reasonable to expect that the worse a state’s fiscal health, the less the risk faced by a public official who supports a tax increase. The evidence using this measure is mixed. Berry and Berry (1990) find a negative but not statistically significant relationship, whereas Erekson et al. (1999) find a negative, statistically significant relationship. It is likely that economic growth in a state is related to the state’s fiscal health. States experiencing a recession may find it especially difficult to increase revenues using conventional forms of taxation, which may lead them to adopt a lottery. Alm, McKee, and Skidmore (1993) find that the larger the percentage change in real state personal income, the less likely is a state to adopt a lottery in a given period. Somewhat surprisingly, however, Alm, McKee, and Skidmore (1993) do not find either the percentage change in state tax revenues or the percentage change in state and local tax revenues to be statistically significant determinants of lottery adoption. One would expect these measures to be more closely tied to fiscal crises and thus to support the fiscal crisis argument. Erekson et al. (1999) use somewhat different measures to capture the percentage change in the tax base in that they use the percentage change in per capita earnings in selected industries. They find some support for the fiscal health argument when they examine growth of earnings in the service sector. Fiscal pressures might be lessened by intergovernmental transfers. Alm, McKee, and Skidmore (1993) examine both the percentage change in intergovernmental transfers to state government only and to state and local government jointly. Such transfers are not statistically significant determinants of lottery adoption. Caudill et al. (1995) and Mixon et al. (1997) genM AY / J U N E 2006 173 Coughlin, Garrett, Hernández-Murillo erate a similar finding using levels rather than percentage changes in intergovernmental transfers. In summary, the fiscal health argument receives, at best, limited empirical support. One additional finding might provide some insights as to why the fiscal health argument applies only weakly to lottery adoption. Fink, Marco, and Rork (2004) found that overall state tax revenues declined with increased lottery sales. This net decline results from a decrease in sales and excise tax revenue, which is offset partially by an increase in income tax revenue. These changes in revenue for specific taxes are related to changes in both economic behavior and tax laws. For example, consumers are likely to substitute, to some degree, the purchase of lottery tickets for the purchase of goods subject to a sales tax. This substitution would increase lottery tax revenue and reduce sales tax revenue. In another study, Fink, Marco, and Rork (2003) find that lottery sales did not have a statistically significant effect on per capita state tax revenues. The implication of these studies is that the adoption of lotteries does not appear to provide even a partial solution to a state’s fiscal problems. Election Cycles and the Political Decision Process Berry and Berry (1990) examine whether the timing of elections might affect the adoption of tax increases. They argue that lotteries will tend to be adopted in election years relative to other years because the lottery relative to other types of tax increases is generally more popular, a fact that elected officials are aware of and likely attempt to use to their advantage.26 Their relative popularity makes lotteries best suited for consideration and adoption in election years. Other taxes, because of their unpopularity, are more likely to be adopted in the year immediately following an election because this provides the maximum time prior to the next election for the electorate to forget about an unpopular tax increase. In addition, for those years that are neither an election year nor the year immediately following an election 26 We are not suggesting that lotteries are uncontroversial, only less controversial as a revenue-increasing option. 174 M AY / J U N E 2006 year, one should expect the probability of lottery adoption to fall somewhere in between—that is, to be less than it is in an election year but more than it is in the year following an election. Berry and Berry (1990) find empirical support for the preceding reasoning; however, Glickman and Painter (2004) do not. The political decision process provides some limited information as to whether a lottery will be adopted in a given period. Alm, McKee, and Skidmore (1993) argue that political pressures for lottery adoption are likely to differ between states that use a referendum or an initiative compared with states that use a standard legislative process. They find that states using either a referendum or an initiative are more likely to adopt a lottery in a given year. Party Control Single-party control of a state’s governorship and both houses of the legislature should make it easier for proposed legislation to be passed and signed into law. An empirical issue is whether such control increases the probability of lottery adoption in a given period. Berry and Berry (1990) hypothesize that such control would be associated with lottery adoption. They found, however, that such control decreased the probability of lottery adoption. One possibility is that a unified government will find it easier to increase existing taxes to achieve substantial revenue increases. Thus, a lottery is not needed to raise revenues. However, in terms of political control, divided governments might find it easier to reach agreement on a lottery, which is a relatively less controversial funding mechanism. The connection between party control and lottery adoption has been explored in a number of other studies, but the results do not provide clear insights into the connection between party control and lottery adoption. For example, Alm, McKee, and Skidmore (1993) explored the role of party control by using separate dummies for Democratic control and Republican control versus shared control. Democratic control was a negative and statistically significant determinant of lottery adoption, whereas Republican control was not statistically significant. Glickman and Painter F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Coughlin, Garrett, Hernández-Murillo Figure 2 Figure 2 Lottery Diffusion by Decade Lottery Diffusion by Decade WA MT ME ND MN OR ID VT WI SD WY MI IA NE NV PA IL UT CA NH MA CT RI NY IN DE DC CO KS NJ OH WV MO VA MD KY AZ NM NC TN OK SC AK MS AL GA TX LA FL AK HI Lottery began in the 1960s Lottery began in the 1990s Lottery began in the 1970s Lottery began in the 2000s Lottery began in the 1980s Non-lottery state SOURCE: Data from Hansen (2004); North Carolina and Oklahoma information from news reports. (2004) find a statistically significant, negative association between the percentage of a state’s lower house that is Democratic and lottery adoption. Finally, Mixon et al. (1997) find no relationship between the percentage of a state’s legislative bodies made up of the majority political party and lottery adoption. Regional Diffusion The spread of lotteries shows a geographic pattern (see Figure 2). Beginning in New Hampshire, lotteries spread to other states in New England, F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W the Mid-Atlantic, and Great Lakes; they then spread throughout states in the Midwest and on the Pacific Coast and began appearing in states in the Plains and Rocky Mountains; most recently, they have spread to the South. In addition to Alaska and Hawaii, the only non-adoptee states are located in the South and the Rocky Mountains. Hansen (2004) notes that the initial reluctance to adopt lotteries stemmed from concerns over the ability of lotteries to raise revenues both efficiently and without corruption. In addition, and more importantly from a geographic perspective, M AY / J U N E 2006 175 Coughlin, Garrett, Hernández-Murillo once lotteries were adopted, cross-border ticket sales were substantial. For example, prior to North Carolina’s passage of a lottery, it was estimated that state residents were spending $100 million per year on Virginia’s lottery.27 Legislators and residents concluded that if residents were going to play the lottery, they would prefer that the spending and resulting tax revenues be kept within their state.28 To account for the possibility that tax competition between state governments might explain lottery adoption, the statistical connection between lottery adoption and a number of geographically based measures have been examined. The results suggest that lottery adoption by a state is related positively to the existence of a lottery in a neighboring state. The only exceptions can be found in Filer, Moak, and Uze (1988) and Glickman and Painter (2004). In both of these studies a dummy variable is used to identify whether an adjacent state had a lottery. On the other hand, this measure is also used by Alm, McKee, and Skidmore (1993), who found a statistically significant, positive relationship. Other studies, such as Davis, Filer, and Moak (1992), Erekson et al. (1999), Caudill et al. (1995), and Mixon et al. (1997), find a statistically significant relationship using the percentage of a state’s border contiguous with states that have lotteries. Caudill et al. (1995) also find that the overall percentage of states already having adopted a lottery tended to increase the probability that a state would adopt a lottery in a given time period. Finally, Berry and Berry (1990) find that a given state was more likely to adopt, the larger the number of adjacent states that had previously adopted. Situational-Specific Determinants In addition to the sets of determinants associated with the adoption of taxes, there are many more constraints and considerations that might influence the adoption of a lottery. For example, 27 See Dube (2005). 28 The importance of cross-border ticket sales, however, does not cease when neighboring states both have lotteries. Tosun and Skidmore (2004) find that the introduction of competing games had an adverse effect on lottery revenues in West Virginia, a state that relies heavily on sales to players in nearby states. 176 M AY / J U N E 2006 Caudill et al. (1995) and Mixon et al. (1997) examine the impact of existing legalized gambling on lottery adoption. They argue that the larger a state’s per capita tax revenue from legalized gambling, the less the need for an alternative revenue source and the more likely the organized opposition to a lottery because of the competitive threat that a lottery poses. Both studies find support for this argument.29 Tax exporting enables the citizens of a state to shift their tax burdens to those outside the state. Generally, taxpayers of a state would prefer to have taxpayers of other states provide the funding for their public services. In terms of lottery adoption, researchers have suggested that the shifting of lottery taxes is easier the larger a state’s tourist industry. Tourists will take advantage of the opportunity to play the lottery and thus will provide lottery revenues. Based on the results of Filer, Moak, and Uze (1988), Davis, Filer, and Moak (1992), Caudill et al. (1995), and Mixon et al. (1997), this argument does not receive empirical support as indices of tax exporting that are based on tourism are not statistically significant determinants of lottery adoption. The decision by a state to adopt a lottery is likely to be related to its prior fiscal decisions. States desiring to raise revenues have various ways to do so; however, some states have fewer alternatives than others. Filer, Moak, and Uze (1988) and Davis, Filer, and Moak (1992) explore the possibility that states without a sales tax are more likely to adopt a lottery than states with a sales tax. Neither study, however, finds a statistically significant relationship. Glickman and Painter (2004) examine the influence of state tax and expenditures limits on lottery adoption and find limits on assessment increases are related to lottery adoption. Finally, Martin and Yandle (1990) examine the impact of state balanced budget requirements and do not find a statistically significant relationship between balanced budget requirements and lottery adoption. 29 In contrast, Davis, Filer, and Moak (1992) find a statistically significant, positive relationship between per capita tax revenue from gambling and lottery adoption. They stress that this finding reflects a preference for gambling. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Coughlin, Garrett, Hernández-Murillo As suggested earlier, the preferences of the electorate are likely to influence political decisions in a democracy. Caudill et al. (1995) and Mixon et al. (1997) use whether a state already has legalized gambling as an indicator of a state’s preferences toward lotteries. Despite the noteworthy exception of Nevada, both studies find that if a state has legalized gambling it is more likely to adopt a lottery. In addition, the preferences of two groups, the elderly and religious groups, have been examined. Alm, McKee, and Skidmore (1993) argue that the elderly tend to oppose most tax increases, but that they might not oppose a form of tax increase that can be viewed as much more voluntary than other forms of tax increases. Using the percentage of a state’s population that is 65 and older, however, neither Alm, McKee, and Skidmore nor Glickman and Painter (2004) find a statistically significant relationship. When one thinks of those in opposition to lottery adoption, one generally starts with religious groups. The lottery battle in Tennessee illustrates this fact. Bobbitt (2003) noted that the largest and most influential anti-lottery group, the Gambling Free Tennessee Alliance, consisted primarily of church groups, such as the Tennessee Baptist Convention, Tennessee Catholic Public Policy Group, and the United Methodist Church. Generally, the most strident opposition by many church groups is based on the belief that gambling is immoral. In addition, other issues such as the prospects of deceptive advertising, the regressivity of the lottery tax, and the prospects for gamblingrelated problems have provided the basis for opposition. With respect to gambling-related problems, the Gambling Free Tennessee Alliance stressed the increased incidence of compulsive gambling and the associated social problems of increased crime, suicide, drug use, and job loss. Various proxies have been used to measure the preferences of religious groups. Frequently used proxies indicating opposition to gambling—used by Filer, Moak, and Uze (1988), Berry and Berry (1990), Martin and Yandle (1990), Caudill et al. (1995), and Mixon et al. (1997)—are either the percentage of a state’s population that is Southern Baptists or, more broadly, that are fundamentalist F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Christians.30 Excluding Filer, Moak, and Uze (1988), these measures are negative and statistically significant determinants of lottery adoption. Two other proxies have also been used. Erekson et al. (1999) use the increase in the percentage of state population that is Protestant. They find a statistically significant, negative relationship with lottery adoption. The final proxy that has been used is the percentage of Catholics in a state. Despite the fact that a Catholic group was part of the Gambling Free Tennessee Alliance, Catholics are often viewed as having a greater preference and tolerance for gambling because of their use of bingo and other games for fundraising purposes. In addition, a larger percentage of Catholics in a state may indicate a smaller percentage of those with religious affiliations who would oppose lotteries. Alm, McKee, and Skidmore (1993) and Glickman and Painter (2004) use such a measure; the former finds a statistically significant, positive relationship with lottery adoption, whereas the latter does not find a statistically significant relationship. Mixon et al. (1997) argue that long-stable societies are more likely to have the special interest groups that provide such stability. These special interest groups—professional associations, labor unions, trade associations, and other coalitions that attempt to shift the distribution of income in their favor—are more likely to exist in older states, which is measured by the years since statehood. Therefore, in light of the redistribution associated with lotteries, the older the state, the more likely a lottery. Mixon et al. find support for this argument. SUMMARY AND CONCLUSIONS Geographic, economic, and political factors have all played roles in the spread of lotteries as well as the decision of some states not to adopt lotteries. On the basis of previous literature, we 30 Fundamentalist Christians encompass those Protestant denominations who believe in the inerrancy of the Bible, the virgin birth of Jesus Christ, the doctrine of substitutionary atonement, the bodily resurrection and return of Christ, and the divinity of Christ. For more details, see “What Is a Fundamentalist Christian?” by Dale A. Robbins at www.victorious.org/chur21.htm. M AY / J U N E 2006 177 Coughlin, Garrett, Hernández-Murillo suggest that explanations for lottery adoption can be organized into six categories: (i) economic development, (ii) fiscal health, (iii) election cycles, (iv) political party control, (v) regional diffusion, and (vi) situational-specific determinants. In terms of economic development, one consistent finding is that higher levels of per capita state income are positively associated with lottery adoption. This finding supports the view that, because those with low incomes bear a higher lottery tax burden than those with high incomes, lotteries are a mechanism for those with high incomes to shift some of their tax burden to those with low incomes. Another explanation is that higher state per capita income is associated with more potential revenue from a lottery. Relatively more urban states have been found to be more likely to have adopted lotteries. A similar comment pertains to more densely populated states. In the former case, the revenue potential argument hinges on the possibility of relatively lower administrative costs in states with relatively larger urban populations. In the latter case, the argument hinges on the possibility of relatively greater values of revenue per dollar of administrative cost due to more potential purchasers of lottery tickets per sales outlet. Although fiscal crises are hypothesized to increase the probability that a state will approve a new tax, a review of existing studies raises doubts about the importance of this explanation for lottery adoption. The various proxies that have been used to capture a state’s fiscal health fail to have a consistent relationship with lottery adoption. One explanation for the absence of a statistically significant relationship between a state’s fiscal health and lottery adoption is that lottery revenues are unlikely to provide sufficient revenues, especially in the near term, to alleviate a fiscal crisis. With regard to political factors that might influence lottery adoption, some empirical support exists for what is termed an election cycle explanation. The political decision to raise taxes is always resisted by at least some individuals and groups; however, the lottery relative to other forms of tax increases is popular. From a politician’s point of view, especially one who desires 178 M AY / J U N E 2006 to be re-elected, the consideration of tax increases in an election year is very risky. As a result, lotteries, which can be viewed as a voluntary tax payment relative to other taxes, are best suited for consideration and adoption in election years. Moreover, even if raising state tax revenues is not an issue, the popularity of lottery adoption might make an election year an especially good time to adopt a lottery. The nature of the decision process does play a role in whether a lottery is ultimately adopted. Relative to a legislative process that relies solely on approval by a state’s legislature prior to a signing by the governor, states that use either an initiative or a referendum as part of the approval process are more likely to have lotteries. The importance of the decision process might be the key to understanding the empirical findings with respect to party control. Findings with respect to party control and lottery adoption fail to provide clear insights concerning the impact of single-party control of a state’s governorship and both houses of the legislature. Such control should make it easier for the controlling party to pass and sign into law its desired legislation. Nonetheless, a consistent empirical relationship between party control and lottery adoption is not identified in the existing studies. Party control might have a lessened effect in states using either an initiative or a referendum because of the direct voicing of the electorate’s preferences. Our review of existing studies highlights the importance of geography in the spread of lotteries across the United States. Regardless of the measure used to account for lotteries in neighboring states, it is clear that lottery adoption by a state is related positively to the existence of lotteries in a neighboring state or states. This finding is a clear indicator of the impact of tax competition between states. State legislators and their constituents have concluded that if residents are going to play the lottery, they would prefer that the spending and tax revenues be kept within their state. It is also reasonable to think that the electorate in a state that already has legalized gambling would be inclined to support lottery adoption. Despite the absence of a lottery in Nevada, empiriF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Coughlin, Garrett, Hernández-Murillo cal support for this line of thinking exists. States with relatively more police and police expenditures are also more likely to have adopted lotteries. Only limited support exists for the argument that prior fiscal decisions that limit a state’s tax revenue potential provide a motivation for lottery adoption. In terms of the preferences of specific groups, both the elderly and religious groups have been examined. For the former group, there is no statistically significant relationship between the percentage of a state’s population that is elderly and lottery adoption. On the other hand, it is clear that the larger the relative size of a religious group that opposes lottery adoption the less likely that a state will adopt a lottery. Religious opposition to lotteries is undoubtedly a key reason that lottery adoption by states in the South tended to lag lottery adoption by states in the rest of the country. One final result suggests that long-stable societies, which are more likely to have entrenched special interest groups, are more likely to have lotteries. Because interest groups attempt to shift the distribution of after-tax income in their favor, lotteries are more likely because of their redistributive effects. This final result is simply one of the many illustrations in this review of lottery adoption showing the connection between economic motivations and political results. Opportunity,” American Journal of Political Science, August 1992, 36(3), pp. 715-42. Blanch, Ernest. “Lotteries and Pools.” American Statistician, 1949, 3(1), pp. 18-21. Bobbitt, Randy. “The Tennessee Lottery Battle: Education Funding vs. Moral Values in the Volunteer State.” Public Relations Quarterly, Winter 2003, 48(4), pp. 39-42. Bonner, Lynn. “Lobbyists Often the Ghost Writers of State Laws.” The News & Observer, October 26, 2005; www.newsobserver.com/656/v-print/story/ 351031.html. Caudill, Stephen B.; Ford, Jon M.; Mixon, Franklin G. and Peng, Ter Chao. “A Discrete-Time Hazard Model of Lottery Adoption.” Applied Economics, June 1995, 27(6), pp. 555-61. Clotfelter, Charles T. and Cook, Philip J. Selling Hope: State Lotteries in America. Cambridge, MA: Harvard University Press, 1989. Clotfelter, Charles T. and Cook, Philip J. “On the Economics of State Lotteries.” Journal of Economic Perspectives, 1990, 4(4), pp. 105-19. REFERENCES Clotfelter, Charles T.; Cook, Philip J.; Edell, Julie A. and Moore, Marian. “State Lotteries at the Turn of the Century: Report to the National Gambling Impact Study Commission.” Report, Duke University, 1999. Alm, James; McKee, Michael and Skidmore, Mark. “Fiscal Pressure, Tax Competition, and the Introduction of Lotteries.” National Tax Journal, December 1993, 46(4), pp. 463-76. Davis, J. Ronnie; Filer, John E. and Moak, Donald L. “The Lottery as an Alternative Source of State Revenue.” Atlantic Economic Journal, June 1992, 20(2), pp. 1-10. Becker, Gary S. “A Theory of Competition among Pressure Groups for Political Influence.” Quarterly Journal of Economics, August 1983, 98(3), pp. 371400. DeBoer, Larry. “Administrative Costs of State Lotteries.” National Tax Journal, December 1985, 38(4), pp. 479-87. Berry, Frances Stokes and Berry, William D. “State Lottery Adoptions as Policy Innovations: An Event History Analysis.” American Political Science Review, June 1990, 84(2), pp. 395-415. Berry, Frances Stokes and Berry, William D. “Tax Innovation in the States: Capitalizing on Political F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Dube, Jonathan. “N.C. Watches as States Add Lotteries,” 2005; www.jondube.com/resume/charlotte/lottery.html. Erekson, O. Homer; Platt, Glenn; Whistler, Christopher and Ziegert, Andrea L. “Factors Influencing the Adoption of State Lotteries.” Applied Economics, July 1999, 31(7), pp. 875-84. M AY / J U N E 2006 179 Coughlin, Garrett, Hernández-Murillo Filer, John E.; Moak, Donald L. and Uze, Barry. “Why Some States Adopt Lotteries and Others Don’t.” Public Finance Quarterly, July 1988, 16(3), pp. 259-83. Fink, Stephen C.; Marco, Alan C. and Rork, Jonathan C. “The Impact of Lotteries on State Tax Revenues.” Proceedings from the Ninety-fifth Annual Conference on Taxation, November 14-16, 2002, Orlando, FL; minutes of the annual meeting of the National Tax Association, November 14, 2002. Volume 27. Washington, DC: National Tax Association, 2003, pp. 1169-72. Fink, Stephen C.; Marco, Alan C. and Rork, Jonathan C. “Lotto Nothing? The Budgetary Impact of State Lotteries.” Applied Economics, December 2004, 36(21), pp. 2357-67. Fisher, Ronald. State and Local Public Finance. Chicago: Irwin, 1996. Garrett, Thomas A. “A Test of Shirking under Legislative and Citizen Vote: The Case of State Lottery Adoption.” Journal of Law and Economics, April 1999, 42(1, Part 1), pp. 189-208. Garrett, Thomas A. “Earmarked Lottery Revenues for Education: A New Test of Fungibility.” Journal of Education Finance, Winter 2001, 26(3), pp. 219-38. Garrett, Thomas A.; Wagner, Gary A. and Wheelock, David C. “A Spatial Analysis of State Banking Regulation.” Papers in Regional Science, November 2005, 84(4), pp. 575-95. Glickman, Mark M. and Painter, Gary D. “Do Tax and Expenditure Limits Lead to State Lotteries? Evidence from the United States: 1970-1992.” Public Finance Review, January 2004, 32(1), pp. 36-64. Hansen, Alicia. “Lotteries and State Fiscal Policy.” Background Paper No. 46, Tax Foundation, October 2004. Hernández-Murillo, Rubén. “Strategic Interaction in Tax Policies Among States.” Federal Reserve Bank of St. Louis Review, May/June 2003, 85(3), pp. 47-56. Hersch, Philip J. and McDougall, Gerald S. “Voting for ‘Sin’ in Kansas.” Public Choice, May 1988, 57(2), pp. 127-39. Martin, Robert and Yandle, Bruce. “State Lotteries as Duopoly Transfer Mechanisms.” Public Choice, March 1990, 64(3), pp. 253-64. McCormick, Robert and Tollison, Robert. Politicians, Legislation, and the Economy: An Inquiry into the Interest Group Theory of Government. Boston: Martinus Nijhoff, 1981. Mixon, Franklin G. Jr.; Caudill, Steven B.; Ford, Jon M. and Peng, Ter Chao. “The Rise (or Fall) of Lottery Adoption within the Logic of Collective Action: Some Empirical Evidence.” Journal of Economics and Finance, Spring 1997, 21(1), pp. 43-49. Oster, Emily. “Are All Lotteries Regressive? Evidence from PowerBall.” National Tax Journal, June 2004, 57(2, Part I), pp. 179-87. Robbins, Dale A. “What Is a Fundamentalist Christian?” 1995; www.victorious.org/chur21.htm. Rodrik, Dani. “Political Economy of Trade Policy,” in Gene M. Grossman and Kenneth Rogoff, eds., Handbook of International Economics. Volume III. Princeton, NJ: Elsevier, 1995, pp. 1457-94. Spindler, Charles. “The Lottery and Education: Robbing Peter to Pay Paul.” Public Budgeting and Finance, Fall 1995, 15(3), pp. 54-62. Tosun, Mehmet Serkan and Skidmore, Mark. “Interstate Competition and State Lottery Revenues.” National Tax Journal, June 2004, 57(2, Part 1), pp. 163-78. Hansen, Susan B. The Politics of Taxation: Revenue without Representation. Westpoint, CT: Praeger, 1983. 180 M AY / J U N E 2006 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W The 1990s Acceleration in Labor Productivity: Causes and Measurement Richard G. Anderson and Kevin L. Kliesen The acceleration of labor productivity growth that began during the mid-1990s is the defining economic event of the past decade. A consensus has arisen among economists that the acceleration was caused by technological innovations that decreased the quality-adjusted prices of semiconductors and related information and communications technology (ICT) products, including digital computers. In sharp contrast to the previous 20 years, services-producing sectors—heavy users of ICT products—led the productivity increase, besting even a robust manufacturing sector. In this article, the authors survey the performance of the services-producing and goods-producing sectors and examine revisions to aggregate labor productivity data of the type commonly discussed by policymakers. The revisions, at times, were large enough to reverse preliminary conclusions regarding productivity growth slowdowns and accelerations. The unanticipated acceleration in the services sector and the large size of revisions to aggregate data combine to shed light on why economists were slow to recognize the productivity acceleration. Federal Reserve Bank of St. Louis Review, May/June 2006, 88(3), pp. 181-202. O ver the past decade, economists have reached a consensus that (i) the trend rate of growth of labor productivity in the U.S. economy increased in the mid-1990s and (ii) the underlying cause of that increase was technological innovations in semiconductor manufacturing that increased the rate of decrease of semiconductor prices.1 This productivity acceleration is remarkable because, unlike most of its predecessors, it continued with only a minor slowdown during the most-recent recession. In this article, we briefly survey research on the genesis of the productivity rebound. We also examine the “recognition problem” that faced economists and policymakers during the 1990s when preliminary data, both economywide and at the industry 1 The first chapter of Jorgenson, Ho, and Stiroh (2005) surveys the decrease in semiconductor prices. level, showed little pick up in productivity growth. Using more than a decade of vintage “real-time” data, we find that revisions to labor productivity data have been large, in some cases so large as to fully reverse initial preliminary conclusions regarding productivity growth slowdowns and accelerations. The 1990s acceleration of labor productivity has three important characteristics. First, it was unforeseen. An example of economists’ typical projections during the mid-1990s is the 1996 Economic Report of the President, prepared during 1995, in which the Council of Economic Advisers foresaw no revolutionary change. The Council foresaw labor productivity growth in the private nonfarm business sector at an average annual rate of 1.2 percent from the third quarter of 1995 to the end of 2002. This estimate largely extrapolated recent experience: productivity from 1973 to 1995 had grown at an average annual rate of 1.4 percent. Richard G. Anderson is an economist and vice president and Kevin L. Kliesen is an economist at the Federal Reserve Bank of St. Louis. The authors thank Aeimit Lakdawala, Tom Pollmann, Giang Ho, and Marcela Williams for research assistance. © 2006, The Federal Reserve Bank of St. Louis. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W M AY / J U N E 2006 181 Anderson and Kliesen Figure 1 Nonfarm Labor Productivity Percent Change from Peak Value, 2001:Q1 to 2005:Q4 20 18 Current Estimate 16 14 Average Business Cycle 12 10 8 6 4 2 0 0 1 2 –2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Quarters after Peak NOTE: The NBER’s business cycle dating committee on November 26, 2001, selected the first quarter of 2001 as the cyclical peak. The business-cycle average is calculated as the mean of the nine NBER post-World War II business cycles, excluding the 1980 and 2001 recessions. Incoming data during 1995 and 1996 were not signaling an increase in productivity growth. Gordon (2002, p. 245) notes that economists in 1997 were still seeking to identify the causes of the post-1973 slowdown in productivity growth: “Those of us who participated in panels on productivity issues at the January 1998 meetings of the American Economic Association recall no such recognition [of a productivity growth rate increase]. Rather, there was singular focus on explaining the long, dismal period of slow productivity growth dating from 1972.”2 Today, with revised data, we know that the productivity acceleration started before 1995. Labor productivity growth showed its resilience by slowing only modestly during the mild 2 As the discussion in Edge, Laubach, and Williams (2004) indicates, Princeton University professor Alan Blinder in 1997 estimated that future, near-term trend labor productivity growth was effectively the same as its average since 1974: 1.1 percent. Further, in 1999, Northwestern University professor Robert Gordon estimated a trend rate of growth of 1.85 percent; he then subsequently revised this up to 2.25 percent in 2000 and then to 2.5 percent in 2003. 182 M AY / J U N E 2006 2001 recession. Forecasters adopted new views of the trend. By 2001, the Council of Economic Advisors had increased its projection of the annual growth of structural labor productivity to 2.3 percent per year. Other forecasters, including many in the Blue Chip Economic Indicators and the Federal Reserve Bank of Philadelphia’s Survey of Professional Forecasters, were even more optimistic.3 Yet, since the March 2001 National Bureau of Economic Research (NBER) business cycle peak, labor productivity has been stronger than both these upward-revised forecasts and its average following past cyclical peaks; the latter point is illustrated in Figure 1. Although no one can be certain of future gains in productivity, it now seems clear that the combination of lower prices for information and communications technology (ICT) equipment plus new related business practices have boosted the economy’s trend rate 3 See the September 10, 2000, issue of the Blue Chip Economic Indicators or the first quarter 2001 Survey of Professional Forecasters. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Anderson and Kliesen of productivity growth. We note that similar increases in labor productivity growth have occurred in other eras and other countries, usually associated with technological innovations.4 Second, the underlying cause was an increase in the rate of decrease of semiconductor prices and, in turn, of ICT capital equipment. In response to falling ICT prices, producers in both servicesproducing and goods-producing sectors shifted increasing amounts of capital investment toward ICT products, reducing in some cases purchases of more traditional capital equipment. Subsequently, many business analysts have noted that, following a gestation lag, the lower cost of ICT equipment has induced firms to “make everything digital” and reorganize their business practices; Friedman (2005) and Cohen and Young (2005) provide detailed case studies. And third, the post-1995 productivity acceleration is largely a services-producing sector story.5 After 1995, productivity growth in services increased sharply while productivity growth in manufacturing continued at approximately its then-extant pace. Ironically, the post-1973 slowdown in aggregate productivity growth also was a services-producing sector story—but one in which productivity in services-producing sectors collapsed.6 Post-1973 pessimists cited Baumol’s 4 Basu et al. (2004) compare and contrast the differing U.S. and U.K. experiences after 1995. 5 In this article, we follow the Bureau of Economic Analysis (BEA) data reporting schema. Before June 2004, the BEA followed the SIC (Standard Industrial Classification) schema. Services-producing industries included transportation and public utilities; wholesale trade; retail trade; finance, insurance, and real estate (FIRE), including depository and nondepository institutions; and services (business services, including computer and data processing services). Private goods-producing industries included agriculture, forestry, and fishing; mining; construction; and manufacturing. In June 2004, the BEA revised its schema to follow the 1997 North American Industry Classification System (NAICS). The composition of services-producing industries changed slightly to include utilities; wholesale trade; retail trade; transportation and warehousing; finance, insurance, real estate, rental and leasing (FIRE); professional and business services, including computer systems design and related services; educational services, health care, and social assistance; arts, entertainment, recreation, accommodation, and food services; and other services, except government. Compared with the SIC, the NAICS more consistently classifies high-technology establishments into the correct industry and provides increased detail on the services sector (Yuskavage and Pho, 2004). We do not, in this analysis, examine interactions between these redefinitions and revisions to published data. 6 See Kozicki (1997). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W (1967) analysis that services sectors had little potential to increase labor productivity and expounded views that the expanding share of services in gross domestic product (GDP) foreshadowed an eternal era of slow labor productivity growth for the U.S. economy.7 As early as the 1973 productivity slowdown, however, the servicesproducing sector was a major user of information technology, poised to benefit from improvements in semiconductor manufacturing. Hence, the significant technological advances in the early 1990s were especially important for services-producing sectors (Triplett and Bosworth, 2004; Jorgenson, Ho, and Stiroh, 2005). The mechanism was straightforward: Sharp decreases in the prices of semiconductors and related ICT capital goods induced services-sector firms to significantly increase their use of ICT capital, in turn increasing productivity growth and, with it, productivity growth for the entire economy. Both then and now, three-quarters of private-sector real GDP arises from services-producing sectors. Poor data quality often has been cited as the barrier to identifying the causes of the post-1973 slowdown in services-sector productivity growth; see Griliches (1992 and 1994) or Sherwood (1994). Measurement issues for services-producing sectors have a long history, largely focused on correct measures of “output,” including the price deflators necessary for obtaining real output from nominal magnitudes. As early as 1983, members of the Federal Open Market Committee (FOMC) questioned the quality of data on output and productivity in services-producing sectors; such discussions became longer and more frequent after Chairman Greenspan’s lengthy soliloquy at the December 22, 1992, meeting.8 In 1996, Chairman Greenspan noted it was implausible that services-sector labor 7 Baumol (1967) argued that some services—including municipal government, education, performing arts, restaurants, and leisure time activities—had a “technological structure” that made longterm increases in the real cost of such services unavoidable because it was unlikely that productivity gains would be large enough to offset increasing wages. Baumol did not suggest, however, that all services-producing sectors were condemned to little or no productivity growth even though some later authors attributed that position to him. 8 Anderson and Kliesen (2005) review the history of productivity discussions in the FOMC transcripts from 1982 to 1999. M AY / J U N E 2006 183 Anderson and Kliesen productivity had not increased during the past 20 years and requested the Board staff to conduct a study of the quality of data for services-producing industries. The resulting study—Corrado and Slifman (1999)—confirmed the problematic quality of services-sector data but concluded that “the output, price, and unit-costs statistics for the nonfinancial corporate sector are internally consistent and economically plausible” (p. 332).9 Yet, even in these data, measured productivity growth in manufacturing was approximately double that in nonmanufacturing: For 1989 to 1997, the increases in output per hour were 2.9 and 1.4 percent for manufacturing and nonmanufacturing, respectively (Corrado and Slifman, 1999, p. 329, Table 1). The situation has improved significantly in recent years. During the past decade, data measurement programs at both the BEA and the Bureau of Labor Statistics (BLS) have produced wellmeasured data for the services sectors, culminating in the BEA’s December 2005 publication of the first NAICS industry-level data fully consistent across their input-output matrices, their annual industry accounts, and the nationwide GDP national income accounts system. Somewhat earlier, resolution of the vexing services-sector productivity problem occurred in 2000 when the BEA incorporated into the annual industry accounts their October 1999 revisions to the national income and product accounts (NIPA).10 9 10 Previously published data had shown some rebound in measured productivity growth for services sectors, but services continued to lag well behind manufacturing. The revised sector and industry data demonstrated that, far from being the laggard, labor productivity growth in servicesproducing sectors had exceeded productivity growth in manufacturing during the 1990s. Two extensive recent analyses are Triplett and Bosworth (2004) and Jorgenson, Ho, and Stiroh (2005). Unfortunately, the studies’ datasets and analytics differ, making direct comparisons of their numerical productivity growth rates difficult.11 For brevity, we cite results from only one of the studies. Triplett and Bosworth (2004) find that labor productivity in services-producing industries increased at an annual average rate of 2.6 percent between 1995 and 2001 (including the 2001 economic slowdown), slightly faster than manufacturing’s 2.3 percent pace. Servicesproducing sectors accounted for 73 percent of 1995-2001 labor productivity growth and 76 percent of multifactor productivity growth (defined below). Increased use of ICT capital was the primary cause behind the productivity acceleration: When weighted by its large share of the economy, increased ICT use in services accounts for 80 percent of the total contribution of ICT to increased economywide labor productivity growth between 1995 and 2001. Their conclusion? On page 2, they write, “As with labor productivity growth and Corrado and Slifman (1999) argued that most data problems were in the nonfinancial noncorporate sector, half of which was composed of difficult-to-measure services-sector firms. They concluded that mismeasurement so contaminated these figures that data for the nonfarm business sector should not be used for analysis. During the 1990s, the BEA greatly expanded and improved its industry database, partly in response to controversy regarding productivity growth. The BEA added gross output (shipments) by industry in 1996 (Yuskavage, 1996). Gross output is more desirable for productivity studies than gross product originating (value added), a point highlighted by Evsey Domar’s much-earlier quip that few people find it interesting to study productivity in shoe manufacturing when leather is omitted. Interested readers can judge the impact of the October 1999 revisions by comparing studies before and after their publication. Such a comparison is not included here because, in our opinion, methodological changes for the annual industry accounts have been so large as to render comparisons of vintage data of questionable value. Typical of the pre-revision analyses is Triplett and Bosworth (2001), a paper originally presented at the January 2000 American Economic Association meetings. Ironically and with more than a touch of understatement, they note that “The nonfarm multifactor productivity numbers are due for revision in the near future, to incorporate the revisions to GDP that 184 M AY / J U N E 2006 were released in October, 1999. This will undoubtedly raise the nongoods estimate but not the manufacturing productivity estimate…” Shortly thereafter, they declared “Baumol’s disease” to be cured; see Triplett and Bosworth (2003 and 2006), the latter paper was originally prepared for an April 2002 conference at Texas A&M University. Interested readers might also compare Gordon (2000 and 2003). One of the earliest studies using the revised data is Stiroh (2002), which first appeared in January 2001 as Federal Reserve Bank of New York Staff Report 115 (the published article contains later, revised data) and showed productivity accelerations in broad service sectors, including wholesale and retail trade and finance, insurance, and real estate. 11 Triplett and Bosworth use output and employment data from the BEA’s annual industry accounts and capital from the BLS’s capital flow accounts. Their labor input measure is persons employed, not hours worked, and is not quality-adjusted. Although these shortcomings perhaps bias upward their estimated level of labor productivity, it seems unlikely that it distorts labor productivity growth significantly over shorter periods (i.e., 5 years or so). Jorgenson, Ho, and Stiroh measure output broadly to include the services of household durable goods and housing. They also use constant-quality index numbers for labor and capital input. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Anderson and Kliesen multifactor productivity growth, the IT [information technology] revolution in the United States is a services industry story.” It is important to note, as they do, that not all services-sector industries had productivity increases; most did, but some services industries continue to have negative measured productivity growth. THE ROLE OF SEMICONDUCTOR PRICES Understanding the sources of the labor productivity acceleration makes it easier to appreciate the difficulties in measuring it. Economists define labor productivity as the ratio of the economy’s real output, Y, to total hours of labor input, H, Y/H. Let us assume that total output is produced by means of an aggregate production function, Y = A × F(H,K), where K measures the flow of productive services from the economy’s capital stock and A measures increases in output not due to increases in labor (H) or capital (K), that is, multifactor productivity (MFP).12 In this framework, there are two sources of increases in labor productivity: capital deepening and increases in MFP. Capital deepening is defined as increases in the amount of capital equipment available per hour worked, K/H. Increases in MFP often are referred to as improvements “in the ways of doing things,” that is, changes in firms’ business management practices. The growth rate of A may be written as, and often is measured as, a residual by means of the equation g A = gY − ⎡⎣(1 − ν ) g K + ν g H ⎤⎦ , where gA, gY, gK , and gH , respectively, are the growth rates of MFP, output, capital services, and labor services and ν is the share of labor in total output. When increases in output, Y, are fully accounted for by increases in H or K by means of the function F (H, K ), then by definition there is no change in A, that is, no increase in MFP (but 12 For simplicity, we are omitting intermediate inputs, making total output equal to value added (real GDP originating). For a richer model that contains intermediate inputs, see Jorgenson, Ho, and Stiroh (2005). The term A also often is referred to as the Solow residual. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W labor productivity may continue to increase through capital deepening). Note that mismeasurement of H or K will cause mismeasurement of MFP. Specifically, if H and K contain unmeasured increases in quality, then measured H and K will tend to understate the flow of labor and capital services and tend to overstate growth of MFP.13 Productivity statistics published by the BLS use, for H, a constant-quality index of labor input that adjusts for such changes by using hourly wage rates to combine the working hours of workers with different characteristics.14 For the brief time periods considered here, labor quality adjustments likely matter little: For 1995-2000, Jorgenson, Ho, and Stiroh (2005) find that the difference between aggregate economywide hours at work and their constant-quality quantity index of labor volume is small. The constant-quality index grew only 0.34 percent faster than total hours worked during the period, largely due to the more rapid growth of higher-paid workers. Measuring capital input, K, has similar issues. The productive capital stock of the United States is a heterogeneous collection of producers’ durable equipment and structures, each with different specific characteristics. Although macroeconomists tend to measure the “capital stock” by summing the constant-dollar purchase prices of all such capital assets (after a depreciation allowance), a superior practice for productivity analysis is to measure K as a constant-quality quantity index of the flow of capital services.15 Jorgenson, Ho, and 13 This relationship has been well-known at least since Solow (1960). Note that Solow (1957) uses a capital stock measure that is not adjusted for quality change and, as a result, captures almost all productivity improvements in A as MFP. Solow (2001) lauds the introduction of constant-quality labor and capital services index numbers. Many macroeconomic studies, however, continue to use capital stock measures unadjusted for quality; see, for example, Jones (2002). 14 Jorgenson, Gollop, and Fraumeni (1987) is the classic study. For a recent discussion that also provides newly constructed measures for ICT-related sectors, see Jorgeonson, Ho, and Stiroh (2005, Chap. 6). 15 Hulten (1992) has a clear exposition of the trade-off between measuring quality change in capital goods and measuring multifactor productivity. Pakko (2002a) explores whether applying reasonable quality adjustments to non-ICT capital investment during the 1990s would change the then-published profile of investment spending. He concludes it would not. M AY / J U N E 2006 185 Anderson and Kliesen Stiroh (2005) emphasize that incorporating quality measures for capital services is essential to understanding the 1990s productivity acceleration. Many analysts have noted the short service lives, rapid depreciation rates, and high marginal products of ICT equipment. Further, the technological innovations that accelerated the fall in semiconductor prices have also allowed the creation of entirely new types of ICT capital goods (as well as innovative consumer goods).16 During the 1990s, for example, businesses shifted their capital investment spending patterns toward relatively shorter-lived ICT capital; information processing equipment and software comprised 25.1 percent of private fixed investment in 2002 versus 11.4 percent in 1977. Both Triplett and Bosworth (2004) and Jorgenson, Ho, and Stiroh (2005) use qualityadjusted capital stock data from the BLS constructed using methods pioneered by Jorgenson. To be specific, a constant-quality quantity index for investment in asset j can be written as Ij = Nominal Investment j PI , j , where PI,j is a constant-quality price index that reflects changes in the productive characteristics and perceived “quality” of the capital asset (time subscripts are omitted). If PI,j is correct, then Ij measures the quantity of new nominal investment in constant-quality “efficiency units” relative to the base year of the price index (Hulten, 1990). Capital stocks are constructed by means of these methods, for example, in Jorgenson, Ho, and Stiroh (2005, Chap. 5). Solow (2001) emphasizes that constant-quality price and quantity indices are subtle concepts, as is the separation of capital deepening from MFP. Fundamentally, all productivity increases are due to increases in knowledge: In some cases, an economist might measure these as increases in quality-adjusted capital and capital deepening; in other cases the economist might identify the 16 Innovations in communications equipment include cell phones, high-density multiplexers for fiber optic cable, and voice-overInternet-protocol (VOIP) telephone equipment. Doms (2005) provides a survey. 186 M AY / J U N E 2006 gains as MFP after quality adjustments have been made to labor and capital inputs. Regardless, increases in the knowledge of how to produce goods and services is the fundamental cause of productivity growth. Quality adjustments often are subjective and uncertain. Judgment errors in the PI,j necessarily affect measures of both capital deepening and MFP. Overestimates of increased quality potentially can inflate constant-quality quantity indices to the extent that MFP vanishes. Solow offers an example to illustrate the issue. Consider a competitive two-sector economy in which one sector produces capital goods from labor (only) and the other produces consumer goods from labor and capital. Let us assume a technological innovation occurs that reduces the quantity of labor required to produce one unit of the capital good but does not change its physical characteristics. In this case, both the observed market price and the constant-quality price index fall (no quality adjustment is made to the observed price of the capital good). As profit-maximizing producers of consumer goods increase the quantity of now less-expensive capital per hour of labor, both the physical capital stock and the constantquality quantity index (as well as the capitallabor ratio) will increase. Alternatively, starting from the same two-sector economy as before, let us assume a technological innovation occurs that increases the productivity of each unit of capital in the production of consumer goods but does not change the amount of labor required to produce each unit of the capital good. In this case, the observed price of the capital good is unchanged but the constant-quality price index for capital goods falls. As profit-maximizing producers of consumer goods replace older capital with newer capital, the constant-quality quantity index for capital will increase even if the physical capital stock does not, and the ratio of constant-quality capital units to labor volume will rise. Under mild assumptions, the long-run equilibrium economic effect of the two alternative technological innovations is exactly the same, although the adjustment may entail a long lag when the technological improvement is largely or entirely embedded in F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Anderson and Kliesen the new capital good (Pakko, 2002b and 2005). The widely used putty-clay model of Greenwood, Hercowitz, and Krussell (1997) is similar to Solow’s example. In their model, a single output good can be “hardened” into consumption goods, investment in structures, or investment in equipment. The production function for output includes labor and both types of capital goods, allowing rich feedback effects among investment-specific technological progress, production of new capital goods, and capital deepening.17 Solow’s example is helpful in understanding how investment in ICT equipment affects productivity. To be specific, it is useful to distinguish between increases in productivity at firms that make high-technology products and at firms that solely use ICT.18 For the former, technological progress in semiconductor manufacturing allows more computing power to be produced from the same inputs of capital and labor because such firms are large users of information technology equipment in development and production. For the latter, decreases in the cost of information technology induce capital deepening—that is, they induce the firm to provide additional capital equipment for each worker. Examples include initiating/expanding e-commerce on the Internet; improving the timeliness of linkages between point-of-sale cash registers and inventory management systems; and improving network links among geographically separated sites. Studies suggest that such changes in business practice may take considerable time to implement; hence, the response of productivity to changes in investment 17 Greenwood, Hercowitz, and Krussell (1997) compare their constantquality capital stock measures for 1954-90, built from data as published circa early 1994, to the capital stock measures calculated by the BEA (which at that time lacked quality adjustment) but not to the constant-quality quantity indices of the BLS. See Dean and Harper (2001) and Jorgenson, Landefeld, and Nordhaus (2006) for comparisons of the BEA and BLS capital measures. 18 Readers are cautioned that Solow’s example, while illustrative, is only an example; recall he assumes that capital goods are made from labor only. In the real world, semiconductor manufacturing is a large user of its own products in the form of computer-assisted design and manufacturing. In the model of Greenwood, Hercowitz, and Krussell (1997), the economy’s output can be hardened into capital which, in following periods, is an input to the production of more output, including future capital goods. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W in ICT equipment varies among firms and industries. Such variation may delay timely recognition that forces have arisen, which, eventually, will increase productivity economywide. EMPIRICAL ESTIMATES OF ACCELERATING AGGREGATE LABOR PRODUCTIVITY The above framework has been used by a number of authors to measure the effect of investment in ICT on productivity growth. Among the more important aggregate (not industry-level) studies are Oliner and Sichel (2002) and Jorgenson, Ho, and Stiroh (2002, 2003, and 2004). Although the studies’ details differ, at the aggregate level of the national economy the authors attribute approximately three-fifths of the acceleration in labor productivity during the second half of the 1990s to capital deepening and two-fifths to increases in MFP. In turn, the authors find that approximately four-fifths of the capital deepening is due to investment in ICT equipment, with increased spending on traditional business equipment accounting for the other one-fifth. Both studies emphasize that purchases of ICT equipment were boosted by rapid decreases in the prices of such equipment, due in large part to rapidly falling prices of component semiconductors, and perhaps displaced to some extent purchases of traditional equipment. As an example of the interaction between measurement and economic modeling, consider the Oliner and Sichel (2002) model. In this model, the rate of increase in MFP is measured by the inverse of the rate of decrease of semiconductor prices, creating a direct link between observed decreases in semiconductor prices and unobserved increases in productivity growth. The intuition is that, because semiconductor prices are falling rapidly relative to the aggregate price level, MFP at semiconductor manufacturers must be increasing; if not, the firms would exit the industry. The effect of this measurement technique is that the sharp decline in semiconductor prices in 1997, shown in Figure 2, appears immediately as an M AY / J U N E 2006 187 Anderson and Kliesen Figure 2 Contributions to Labor Productivity Growth and Relative Changes in Semiconductor Prices Percent Contribution from Capital Deepening 4 Contribution from MFP Percent Change in Relative Semiconductor Prices (right axis) 3 Percent 10 0 –10 –20 2 –30 1 –40 –50 0 –60 –1 –70 –2 1983 1986 1989 1992 1995 1998 2001 2004 –80 SOURCE: Productivity data, Dan Sichel (via e-mail); semiconductor prices, BLS. increase in labor productivity growth. More recent estimates provided to the authors by Dan Sichel, shown in Figure 2, suggest that the direct contribution from the semiconductor industry was responsible for 0.08 percentage points of the 0.37 percent growth of MFP from 1974 to 1990 and 0.13 percentage points of the 0.58 percent growth from 1991 to 1995; after 1995, the proportions change.19 He estimates that the direct contribution from the semiconductor industry from 1996 to 2003 was responsible for 0.40 percentage points of the economy’s total 1.34 percent annual growth of MFP. Complementary analyses are presented by Jorgenson, Ho, and Stiroh (2002 and 2004). (The latter paper’s results differ from the former’s because of data revisions.) In their 2004 analysis, labor productivity (adjusted for shifts in labor quality) increased during the 1995-2003 period at a rate 1.6 percentage points greater than during the 1973-95 period; they attribute a little less than three-fifths of this increase to capital deepening. If the acceleration of productivity was driven by an increase in the rate of decrease of semicon19 Unpublished estimates received from Dan Sichel via e-mail correspondence on June 28, 2004. 188 M AY / J U N E 2006 ductor (and computer) prices, just how fast did prices fall? As shown in Figure 2, semiconductor prices decreased throughout the 1990s with the rate of decrease accelerating during the latter half of the decade. Caution must be used in interpreting these figures, however, because rapid technological change has introduced thorny quality-adjustment problems. The caution expressed by Gullickson and Harper (2002) is typical: These findings rest on estimated trends for high tech inputs and outputs that incorporate adjustments to account for changes in their quality. Many of the high tech input and output growth rates are well up in the double-digit percentage range. These extraordinary trends, in turn, rest on the use of quality adjusted price indexes in deflation. These indicate that prices for high tech goods of constant quality have fallen very rapidly. These price trend estimates have withstood much scrutiny, but we must emphasize their importance for our conclusions. While it is likely that real output trends have been underestimated in many or all of the service sector industries with negative MFP trends, it is also possible that the growth trends for high tech inputs have been overestimated. Underestimating service sector output trends would bias the aggregate productivity trend downF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Anderson and Kliesen ward. Overestimating high tech input and output trends would bias the aggregate productivity trend upward...We can express a concern that the “measurement playing field” may not be level. We have very intricate means of making quality adjustments to high tech goods, but we have few means to make quality adjustments to service outputs. In other cases, the survey sample for some products, such as semiconductors, has changed.20 Holdway (2001, p. 15), cautions: It would be disingenuous to imply that the PPI has been able to properly value and account for technological change in its cmpu [CPU] price measurements. The standard PPI methodologies for valuing quality change [are] rather limited when faced with quality improvements that are accompanied by reduced input costs due to shifts in the production function. Holdway also notes that the apparent acceleration of semiconductor price decreases during early 1997, as shown in Figure 2, most likely is a result of the introduction of secondary-source pricing data.21 Interested readers also should see Grimm (1998) and Landefeld and Grimm (2000). Since 2000, the relative price of quality-adjusted semiconductors (and related products) has decreased at a slower rate than during the latter part of the 1990s; see Figure 2. Even though the relative prices of semiconductors fell by approximately 38 percent in 2004, this was less than its average decline of approximately 65 percent from 1998 to 2000. 20 For semiconductor prices, for example, the BLS has a series in the producer price index, the BEA has a series used in the national income accounts, and the Federal Reserve Board has a price measure used in its industrial production index. See, for example, Hulten (2001). The semiconductor price series plotted in Figure 7 is the PPI measure relative to the GDP price index. 21 Secondary source prices are price figures collected from catalogs and industry publications, rather than from the manufacturer’s price list. Holdway doesn’t speculate on whether secondary-source price data, if available, might change the pre-1997 trend, but the absence of such data introduces a risk into any study that attributes the productivity acceleration to more rapid price decreases: Would the studies reach the same conclusion if the rate of price decrease from 1993 to 1997 had been the same as that beginning in 1997? Or did the decision to solicit secondary-source price data reflect observations of increased pricing pressure? F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W LABOR PRODUCTIVITY: MEASUREMENT, VOLATILITY, AND REVISIONS Published labor productivity growth rates have two characteristics that complicate recognizing changes in trend growth: volatility and revisions. Volatility is illustrated in Figure 3, which shows compound annual growth rates calculated from the most recently published data for 1-, 4-, and 40-quarter intervals. The high volatility is obvious. Beyond volatility, the figure also illustrates that “trend” labor productivity growth since World War II appears to have gone through three phases: more rapid growth from 1948 to 1973; slower growth from 1973 to 1994; and more rapid growth beginning circa 1995. Measured labor productivity growth in the nonfarm business sector, for example, averaged 3 percent per annum during 1949 to 1972 but less than half this pace during 1973 to 1994, despite strong productivity growth in manufacturing. Since 1995, the pace of productivity growth in the total nonfarm business sector has been about equal to its rate during the earlier high-growth period of 1949 to 1972; for the larger total private business sector, growth over the past 10 years still remains modestly below its earlier pace. The lower two sections of Table 1 decompose productivity growth into growth of its numerator (output) and of its denominator (hours). The increase in productivity growth from 1973-94 (column 2) to 19952004 (column 3) reflects both more rapid growth of the numerator (output) and slower growth of the denominator (hours). For broad sectors, the table shows that the post-1973 productivity growth slowdown (compare columns 1 and 2) largely was due to slowdowns in the services and nondurable manufacturing sectors—durable manufacturing’s labor productivity growth increased modestly throughout the slowdown period. During the most recent decade, durable manufacturing’s productivity growth has jumped to an average annual pace of approximately 5.75 percent, double its 1949-72 pace. Published measurements of the economy’s output and labor input are frequently revised. Not only do data revisions complicate the task facing M AY / J U N E 2006 189 Anderson and Kliesen Figure 3 Labor Productivity Growth, Nonfarm Business Sector 10-Year Growth Rate Compound Annual Rate, Percent, Quarterly Data 3.5 3 2.5 2 1.5 1 0.5 0 1947 1950 1953 1956 1959 1962 1965 1968 1971 1974 1977 1980 1983 1986 1989 1992 1995 1998 2001 2004 1-Year Growth Rate Compound Annual Rate, Percent, Quarterly Data 8 6 4 2 0 –2 –4 1947 1950 1953 1956 1959 1962 1965 1968 1971 1974 1977 1980 1983 1986 1989 1992 1995 1998 2001 2004 1-Quarter Growth Rate Compound Annual Rate, Percent, Quarterly Data 20 15 10 5 0 –5 –10 –15 1947 1950 1953 1956 1959 1962 1965 1968 1971 1974 1977 1980 1983 1986 1989 1992 1995 1998 2001 2004 190 M AY / J U N E 2006 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Anderson and Kliesen Table 1 Decomposition of Average Labor Productivity Growth for the Business Sector Growth for periods indicated Output per hour Business Nonfarm business Manufacturing Durable Nondurable Nonfinancial corporate business Output Business Nonfarm business Manufacturing Durable Nondurable Nonfinancial corporate business Hours Business Nonfarm business Manufacturing Durable Nondurable Nonfinancial corporate business 1949-72 1973-94 1995-2005 1949-2005 3.23 2.77 2.58 2.64 2.83 2.61 1.58 1.48 2.59 3.02 1.90 1.40 2.76 2.69 4.44 5.86 2.85 3.34 2.49 2.25 2.94 3.40 2.47 2.21 4.10 4.22 3.74 4.21 3.48 5.51 3.18 3.17 2.51 2.87 1.90 3.23 3.61 3.64 2.38 4.19 0.16 4.27 3.65 3.70 3.00 3.68 2.22 4.17 0.84 1.41 1.14 1.53 0.63 2.86 1.57 1.66 –0.08 –0.15 0.00 1.81 0.83 0.92 –1.97 –1.58 –2.61 0.90 1.12 1.41 0.05 0.27 –0.25 1.92 NOTE: Compounded annual growth rates using quarterly data: 1949:Q1 to 1972:Q4; 1972:Q4 to 1994:Q4; 1994:Q4 to 2005:Q4. Data for nonfinancial corporations begins in 1958:Q1 and ends in 2005:Q3. Data for total manufacturing and durable and nondurable manufacturing are on an SIC basis prior to 1987. Data for total manufacturing and durable and nondurable manufacturing are on an SIC basis prior to 1987. SOURCE: BLS. policymakers—changing perceived strength or weakness of economic conditions that inform their judgments—but they are often significant enough to dramatically alter economic history.22 As an example, each year the BEA revises the national income and product accounts and the BLS revises employment and aggregate hours worked in the establishment survey. Selected revisions, and their effects, are shown in Table 2. For 1998 and 1999, for example, measured output growth in the sub22 See Himmelberg et al. (2004), Kozicki (2004), Orphanides and van Norden (2005), or Runkle (1998). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W sequent, revised data is sharply higher than in the earlier, preliminary data.23 Beginning with 2001, however, the pattern changed: Measured output growth in the revised data has tended to be lower than in previous, preliminary figures. The NIPA revisions published during mid-2005, for example, trimmed measured real GDP growth over the previous three years by 0.3 percentage points per year, to approximately 3.25 percent. 23 The 1999 revisions, it should be noted, were boosted by the reclassification of software purchased by businesses as fixed investment, rather than as an intermediate expense; see Gullickson and Harper (2002). M AY / J U N E 2006 191 Anderson and Kliesen Table 2 Major Statistical Revisions Since 1996 and Real-Time Estimates of Their Effects Statistical series Major aspects of revision Estimated magnitude of revision January 1996 Comprehensive revision of the NIPA Switch to chain-weighted price indices from fixed-weighted price indices in the NIPA. Government investment defined differently. New methodology for calculating depreciation of fixed capital. Revised estimates show real GDP grew at a 3.2 percent annual rate from 1959 to 1984, 0.2 percentage points faster than old estimate. Real GDP growth from 1987 to 1994 was lowered 0.1 percentage point.* July 1998 Annual revision of the NIPA Updated source data. Methodology changes to expenditures and prices for autos and trucks; improved estimates for several categories of consumer expenditures for services; new method of calculating change in business inventories; some purchases of software by businesses classified as expenses (removed from business fixed investment). From 1994:Q4 to 1998:Q1 the growth of real GDP was revised 0.3 percentage points higher to 3.4 percent; growth of real fixed investment revised 0.6 percentage points higher to 12.7 percent; growth of GDP price index reduced 0.3 percentage points to 1.8 percent. February 1999 Consumer price index (CPI) Switch to geometric means estimation to eliminate lowerlevel bias; affected 61 percent of consumer expenditures. According to the BLS, this switch will reduce the annual rate of increase of the CPI by 0.2 percentage points per year. According to the CEA, methodological changes to the CPI from 1994 to 1999 reduced the annual rate of increase of the CPI by 0.6 percentage points in 1999 compared with the 1994 estimate.† October 1999 Comprehensive revision of the NIPA Introduction of CPI geometric weights; classification of software as a fixed investment; incorporated data from the latest 5-year economic census and 1992 benchmark input-output accounts. From 1987 to 1998, these revisions boosted the annual rate of growth of real GDP by an average of 0.4 percentage points per year.‡ July 2001 Annual revision of the NIPA Updated source data (for example, Census Bureau Annual Surveys); new price index for communications equipment from Federal Reserve Board; monthly data used to calculate GDP converted from SIC to NAICS. Growth of real GDP during revision period (1998:Q1 to 2001:Q1) reduced from 4.1 percent to 3.8 percent (compared with pre-revision estimates). Publication date 192 M AY / J U N E 2006 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Anderson and Kliesen Table 2, cont’d Major Statistical Revisions Since 1996 and Real-Time Estimates of Their Effects Publication date Statistical series Major aspects of revision Estimated magnitude of revision July 2002 Annual revision to the NIPA Updated source data (for example, Census Bureau annual surveys); new methodology for estimating quarterly wages and salaries; new price index within PCE services. Growth of real GDP during revision period (1999:Q1 to 2002:Q1) reduced from 2.8 percent to 2.4 percent (compared with pre-revision estimates). July 2004 Annual revision to the NIPA Update source data; only minor changes in methodology for treatment of health care plans for retired military and measurement of motor vehicle inventories. Growth of real GDP during revision period (2000:Q4 to 2004:Q1) was unchanged at 2.5 percent; growth of real fixed investment in equipment and software revised 0.6 percentage points lower. July 2005 Annual revision to the NIPA Updated source data; incorporation of Census’ quarterly services survey for investment in computer software and for consumer spending for services; improved method of calculating implicit services provided by commercial banks. BEA claims these changes will reduce the volatility of the price index for PCE. Growth of real GDP from 2001:Q4 to 2005:Q1 reduced from 3.5 percent to 3.2 percent. Over the same period, growth of GPD price index and the core PCE price index were revised 0.2 percentage points higher to 2.2 and 1.7 percent, respectively. NOTE: Discussion and estimates of annual revisions to the NIPA were taken from archived reports at their web site: www.bea.gov. SOURCE: *1996 Economic Report of the President, p. 48. †2000 Economic Report of the President, p. 61. ‡Ibid, p. 81. Revisions to national income data change measured productivity, often significantly. Changes since 1994 are summarized in Table 3.24 Consistent with revisions to output, in both 1998 and 1999 the BLS revised upward measured nonfarm labor productivity, and in 2001 and 2002 it revised downward measured productivity. The 2001 revision, for example, reduced the measured three-year growth rate of labor productivity by more than three-quarters of a percentage point. Overall, revisions to productivity growth primarily are due to revisions to measured output and not to revisions in measured employment or aggregate 24 These revisions incorporate both the annual three-year revisions to the NIPA as well as the periodic comprehensive revisions, which occur about every five years. See the footnote to Table 3. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W hours worked. Since 1994, for example, the mean absolute revision to the growth rate of output, 0.30 percentage points, is more than double that of hours worked, 0.14 percentage points, and approximately equal to that of the growth rate of productivity growth, 0.28 percentage points.25 A longer-horizon picture of historical revisions to measured labor productivity growth is shown in Figure 4. For each year, 1959 to 2004, the figure has one vertical line that summarizes all the values of that year’s labor growth as published in various issues of the Economic Report of the President. 25 The BLS’s annual benchmark revisions to establishment data have become smaller over time. From 1984 to 2004, the absolute percentage change in nonfarm payrolls averaged 0.2 percent, a third as much as the 1964-83 period. See Haltom, Mitchell, and Tallman (2005). M AY / J U N E 2006 193 Anderson and Kliesen Table 3 Effect of Annual NIPA Revisions on Measured Growth of Labor Productivity, Output, and Hours in the Nonfarm Business Sector (percent change at a compound rate) Output per hour Output Hours NIPA revision period Initial Revised Difference Initial Revised Difference Initial Revised Difference 1994 2.55 2.36 –0.19 3.99 4.12 0.13 1.40 1.70 0.30 1995 1.72 1.68 –0.04 4.70 4.85 0.15 2.87 3.09 0.22 1996 0.83 0.57 –0.26 2.92 2.84 -0.08 2.03 2.28 0.25 1997 0.75 0.88 0.13 3.21 3.33 0.12 2.40 2.44 0.04 1998 1.55 2.06 0.51 4.15 4.76 0.61 2.60 2.64 0.04 1999 2.31 2.60 0.29 4.44 4.74 0.30 2.38 2.42 0.04 2000 3.30 3.30 0.00 5.41 5.51 0.10 2.06 2.13 0.07 2001 3.05 2.28 –0.77 4.28 3.60 –0.68 1.16 1.27 0.11 2002 3.08 2.71 –0.37 2.00 1.44 –0.56 –1.08 –1.27 –0.19 2003 2.87 3.60 0.73 1.56 1.50 –0.06 –1.27 –2.02 –0.75 2004 4.69 4.45 –0.24 4.33 4.23 –0.10 –0.38 –0.23 0.15 2005 3.97 3.68 –0.29 4.76 4.59 0.17 0.76 0.87 0.11 Mean revision –0.04 0.01 0.03 Mean absolute revision 0.32 0.26 0.19 NOTE: Pre- and post-benchmark figures as published in the BLS Productivity and Cost Report. The NIPA revision period is the nine quarters up to and including the first quarter of the year indicated. The year indicated is the year of publication of the NIPA revision, usually July or August. The 1999 NIPA revision, more extensive than most, incorporated the October 28, 1999, introduction of computer software into business fixed investment. This resulted in revisions back to 1959. Nevertheless, for consistency, the revisions shown here are for the nine quarters ending in the first quarter of the year indicated. (The 1999 revisions to “hours” appeared in the August 5, 1999, Productivity and Cost Report.) The lower and upper ends of each line correspond to the lowest and highest published growth rates, respectively, for that year, while the “dot” indicates the most recent estimate. For many years, the minimum-to-maximum range equals or exceeds 2 percentage points. Ranges for years after 1995 are smaller, perhaps due to better measurement techniques, or perhaps because there are fewer observations. Further insight can be gained from “case studies” of periods during which breaks in trend productivity growth occurred. Here, we consider 1973 and 1995-96. • For 1973, the first-published estimate of labor productivity growth was approximately 3 percent; see Figure 5. This value 194 M AY / J U N E 2006 fell sharply in subsequent revisions. During the late 1980s, however, the published value began to increase. In the most recently published data, 1973’s measured productivity growth is greater than its initially published value—removing entirely any “slowdown” during the year.26 • For 1995 and 1996, the most recently published values differ sharply from initial estimates. For 1995, the most recent value is much lower than the initial estimate; see Figure 6. For 1996, the most recent figure is much higher than the initial estimate; see 26 In this vein, it appears that the switch to chain weights from fixed weights in 1996 (see Table 2) was particularly significant. See Gullickson and Harper (2002). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Anderson and Kliesen Figure 4 Revisions to Real-Time Estimates of Labor Productivity Growth, 1959-2004 1959-69 1970-79 Percentage Points 6.0 Percentage Points 6.0 5.0 5.0 4.0 4.5 4.0 3.8 3.6 3.5 3.1 3.0 3.4 3.1 3.0 4.0 3.3 3.0 2.0 1.6 1.3 0.0 1.7 1.2 1.0 2.7 1.5 1.0 2.0 3.3 3.1 –0.3 –1.0 0.1 0.0 Current Estimate –1.5 –2.0 –3.0 Current Estimate 1980-89 78 79 19 77 19 19 76 19 75 19 73 19 74 72 19 19 19 19 70 63 19 64 19 65 19 66 19 67 19 68 19 69 62 19 61 19 59 60 19 19 19 71 –4.0 –1.0 1990-2004 Percentage Points Percentage Points 5.0 6.0 4.5 4.0 5.0 3.1 3.0 2.0 1.7 1.5 1.4 –0.2 1.9 1.7 –1.0 0.4 –1.0 Current Estimate –2.0 2.7 3.8 3.4 2.5 1.6 1.2 1.0 0.0 –1.0 2.8 2.8 2.7 2.0 0.7 0.5 0.0 4.0 3.0 2.0 1.0 4.1 4.0 0.5 Current Estimate 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 19 91 90 19 19 88 89 19 87 19 19 86 85 19 19 19 84 83 19 82 19 81 19 19 80 –2.0 SOURCE: Economic Report of the President, annual issues, 1959-2004. Figure 7. The revision patterns for 1995 and 1996 made it difficult to recognize, during 1995 and 1996, that a change in trend productivity growth was occurring. Although the initially published estimates for the first three quarters of 1995 suggested a productivity acceleration, by mid-1996 these estimates had been revised downward to less than 1 percent. For 1996, initial estimates for all four quarters were between approximately 0.5 and 1.5 percent, hardly supportive of acceleration. Not until the third quarter of 1997 did revised estimates sugF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W gest an acceleration, and not until mid-1998 was its extent clearly visible in the revised data. Differences between first-published and most recently published productivity figures for 1985 to 2005 are summarized in Table 4 and Figures 8 and 9. The principal conclusion to be drawn from Table 4 is that, although mean revisions are small, mean absolute revisions are large, in some cases approximately equal to the estimated annual growth rate itself. Revisions to four-quarter growth rates are smaller than revisions to one-quarter growth rates, although this is due, in part, to the M AY / J U N E 2006 195 Anderson and Kliesen Figure 5 Real-Time Estimates of 1973 Labor Productivity Growth Percent 3.50 3.00 2.50 2.00 1.50 1.00 1974 1977 1980 1983 1986 1989 1992 1995 1998 2001 2004 Publication Date of Economic Report of the President Figure 6 Labor Productivity Growth, 1995 Year-Over-Year Percent Change, Quarterly; Monthly Figures, Jan 1995–Dec 2000 4 3.5 3 1995:Q1 1995:Q2 1995:Q3 1995:Q4 2.5 2 1.5 1 0.5 0 –0.5 –1 Jan Apr Jul Oct Jan Apr Jul Oct Jan Apr Jul Oct Jan Apr Jul Oct Jan Apr Jul Oct Jan Apr Jul Oct 95 95 95 95 96 96 96 96 97 97 97 97 98 98 98 98 99 99 99 99 00 00 00 00 196 M AY / J U N E 2006 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Anderson and Kliesen Figure 7 Labor Productivity Growth, 1996 Year-Over-Year Percent Change, Quarterly; Monthly Figures, Jan 1996–Dec 2000 4.00 3.50 3.00 2.50 2.00 1.50 1996:Q1 1996:Q2 1996:Q3 1996:Q4 1.00 0.50 0.00 Jan Apr Jul Oct Jan Apr Jul Oct Jan Apr Jul Oct Jan Apr Jul Oct Jan Apr Jul Oct 96 96 96 96 97 97 97 97 98 98 98 98 99 99 99 99 00 00 00 00 Figure 8 Nonfarm Business Sector Labor Productivity Growth Estimates (four-quarter growth rate) Published Value as of 2005 7 6 5 4 3 2 y = 0.4917x + 1.2572 R2 = 0.331 1 –2 0 0 –1 1 2 3 4 5 6 7 Initial Published Value –1 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W M AY / J U N E 2006 197 Anderson and Kliesen Figure 9 Nonfinancial Corporate Business Labor Productivity Growth Estimates (four-quarter growth rate) Published Value as of 2005 10 8 6 4 y = 0.5566x + 1.0938 R2 = 0.5079 2 0 –4 –6 –2 0 2 4 6 8 10 Initial Published Value –2 –4 –6 arithmetic of expressing all changes—including those for one quarter—as annualized growth rates. Note that revisions to output growth rates are smaller than those for productivity and that revisions to hours worked are smaller than revisions to output—suggesting that hours worked may be measured, at least in the near-term, with less error than output. Among the aggregate business sectors, durable goods manufacturing has the largest mean absolute revision. The larger revision likely reflects the better near-term precision with which this sector is measured, including more timely incoming revised data. Two similar conclusions are suggested by Figures 8 and 9. First, there are large differences between first-published data and revised data. Second, more-accurate measurement matters: Revisions for the narrower and somewhat bettermeasured nonfinancial corporate business sector are smaller than for the broader and less wellmeasured nonfarm private business sector. 198 M AY / J U N E 2006 CONCLUSIONS Since 1995, estimates of the economy’s longrun, or structural, rate of labor productivity growth have increased significantly. After having increased at about a 1.4 percent annual rate from 1973 to 1994, the current sustainable pace of labor productivity growth in the nonfarm business sector is widely believed to be from one-half to 1 percentage point higher. Recognition during the mid-1990s of the acceleration of productivity was delayed by weaknesses in measuring productivity. Initial aggregate data for 1995 and 1996, for example, showed little increase in measured productivity. Although these productivity measurements were at odds with both anecdotal observations at individual firms and available data on business investment spending (which suggested that rapidly falling semiconductor and computer prices were encouraging significant capital deepening), not until mid-1997 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Anderson and Kliesen Table 4 Initially Published vs. Most Recently Published Growth Rates of Nonfarm Labor Productivity, 1985:Q3 to 2005:Q4 Output per hour Mean revision Mean absolute revision Output Mean revision Hours Mean absolute revision Mean revision Mean absolute revision 1.55 –0.02 1.07 Growth from preceding period (quarterly, percent annual rate) Business sector Nonfarm Manufacturing Durable Nondurable Nonfinancial corporate 0.40 1.78 0.37 0.41 1.80 0.34 1.54 –0.07 1.09 0.03 2.12 –0.06 1.65 –0.06 1.28 –0.08 2.77 –0.04 2.21 0.05 1.42 0.08 2.16 –0.18 1.73 –0.22 1.38 –0.12 2.01 –0.12 1.82 –0.02 1.04 Growth from corresponding period 1 year earlier (quarterly, percent annual rate) Business sector Nonfarm Manufacturing 0.26 1.03 0.24 0.81 –0.02 0.62 0.25 1.00 0.19 0.80 –0.05 0.62 –0.16 1.37 –0.18 0.89 –0.01 0.77 Durable –0.28 1.89 –0.19 1.43 0.11 0.83 Nondurable –0.11 1.13 –0.32 0.71 –0.18 0.79 –0.04 1.12 –0.04 1.04 –0.02 0.70 Nonfinancial corporate NOTE: Each figure is equal to the initially published growth rate minus the most recently published growth rate for the span indicated. SOURCE: BLS, Productivity and Cost. did revised data for 1995 and 1996 display gains in productivity growth. Our analysis suggests that such measurement delays and revisions are not uncommon. REFERENCES Anderson, Richard G. and Kliesen, Kevin L. “Productivity Measurement and Monetary Policymaking during the 1990s.” Working Paper 2005-067, Federal Reserve Bank of St. Louis, October 2005. Basu, Susanto; Fenald, John G.; Oulton, Nicholas and Srinivasan, Sylaja. “The Case of the Missing Productivity Growth: Or, Does Information Technology Explain Why Productivity Accelerated in the US but Not in the UK?” NBER Working Paper No. 10010, National Bureau of Economic Research, F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W October 2003, published in Mark Gertler and Kenneth Rogoff, eds, NBER Macroeconomics Annual, 2003. Cambridge, MA: MIT Press, 2004. Baumol, William J. “Macroeconomics of Unbalanced Growth: The Anatomy of Urban Crisis.” American Economic Review, June 1967, 57(3), pp. 415-26. Cohen, Linda and Young, Allie. Multisourcing: Moving Beyond Outsourcing to Achieve Growth and Agility. Cambridge, MA: Harvard Business School Press, 2005. Corrado, Carol and Lawrence, Slifman. “Decomposition of Productivity and Unit Costs.” American Economic Review, May 1999, 89(2), pp. 328-32. Dean, Edwin R. and Harper, Michael J. “The BLS Productivity Measurement Program,” in Charles R. Hulten, Edwin R. Dean, and Michael J. Harper, eds., New Developments in Productivity Analysis. M AY / J U N E 2006 199 Anderson and Kliesen Chicago: University of Chicago Press for the NBER, 2001, pp. 55-84. Doms, Mark. “Communications Equipment: What Has Happened to Prices?” in Carol Corrado, John Haltiwanger, and Daniel Sichel, eds., Measuring Capital in the New Economy. Chicago: University of Chicago Press for the NBER, 2005, pp. 323-62. Edge, Rochelle M.; Laubach, Thomas and Williams, John C. “Learning and Shifts in Long-Run Productivity Growth.” Working Paper 2004-21, Board of Governors of the Federal Reserve System, April 2004. Friedman, Thomas L. The World Is Flat. New York: Farrar, Straus, and Giroux, 2005. Ellis W. “Payroll Employment Data: Measuring the Effects of Annual Revisions.” Federal Reserve Bank of Atlanta Economic Review, Second Quarter 2005, pp. 1-23. Himmelberg, Charles P.; Mahoney, James M.; Bang, April and Chernoff, Brian. “Recent Revisions to Corporate Profits: What We Know and When We Knew It.” Federal Reserve Bank of New York Current Issues in Economics and Finance, March 2004, 10(3), pp. 1-7. Holdway, Mike. “An Alternative Methodology: Valuing Quality Change for Microprocessors in the PPI.” Presentation to the Advisory Committee to the Bureau of Economic Analysis, May 11, 2001. Gordon, Robert J. “Does the ‘New Economy’ Measure Up to the Great Inventions of the Past?” Journal of Economic Perspectives, 2000, 14(4), pp. 49-74. Hulten, Charles R. “The Measurement of Capital,” in Ernst R. Berndt and Jack E. Triplett, eds., Fifty Years of Economic Measurement. Chicago: University of Chicago Press for the NBER, 1990, pp. 119-52. Gordon, Robert J. “Productivity Growth and the New Economy: Comments.” Brookings Papers on Economic Activity, 2002, 0(2), pp. 245-53. Hulten, Charles R. “Growth Accounting When Technical Change Is Embodied in Capital.” American Economic Review, September 1992, 82(4), pp. 964-80. Gordon, Robert J. “Exploding Productivity Growth: Context, Causes and Implications.” Brookings Papers on Economic Activity, 2003, 0(2), pp. 207-98. Hulten, Charles R. “Total Factor Productivity: A Short Biography,” in Charles R. Hulten, Edwin R. Dean, and Michael J. Harper, eds., New Developments in Productivity Analysis. Chicago: University of Chicago Press for the NBER, 2001 Greenwood, Jeremy; Hercowitz, Zvi and Krusell, Per. “Long-Run Implications of Investment-Specific Technological Change.” American Economic Review, 1997, 87(3), pp. 342-62. Griliches, Zvi. Output Measurement in the Service Sectors. Chicago: University of Chicago Press, 1992. Griliches, Zvi. “Productivity, R&D, and the Data Constraint.” American Economic Review, March 1994, 84(1), pp. 1-23. Grimm, Bruce T. “Price Indexes for Selected Semiconductors 1974–96.” Survey of Current Business, February 1998, 78(2), pp. 8-24. Gullickson, William and Harper, Michael J. “Bias in Aggregate Productivity Trends Revisited.” Monthly Labor Review, March 2002, 125(3), pp. 32-40. Haltom, Nicholas L.; Mitchell, Vanessa D. and Tallman, 200 M AY / J U N E 2006 Jones, Charles I. “Sources of U.S. Economic Growth in a World of Ideas.” American Economic Review, March 2002, 92(1), pp. 220-39. Jorgenson, Dale W.; Gollop, Frank M. and Fraumeni, Barbara M. Productivity and U.S. Economic Growth. Harvard Economic Studies. Volume 159. Cambridge, MA: Harvard University Press, 1987. Jorgenson, Dale W.; Ho, Mun S. and Stiroh, Kevin J. “Projecting Productivity Growth: Lessons from the U.S. Growth Resurgence.” Federal Reserve Bank of Atlanta Economic Review, Third Quarter 2002, pp. 1-13. Jorgenson, Dale W.; Ho, Mun S. and Stiroh, Kevin J. “Lessons from the U.S. Growth Resurgence.” Paper prepared for the First International Conference on the Economic and Social Implications of Information F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Anderson and Kliesen Technology, Washington, DC, January 27-28, 2003; www.si.umich.edu/~kahin/hawk/htdocs/ jorgensonpaper.doc. Jorgenson, Dale W.; Ho, Mun S. and Stiroh, Kevin J. “Will the U.S. Productivity Resurgence Continue?” Federal Reserve Bank of New York Current Issues in Economics and Finance, December 2004, 10(13). Jorgenson, Dale W.; Ho, Mun S. and Stiroh, Kevin J. “Projecting Productivity Growth: Lessons from the U.S. Growth Resurgence,” in William H. Dutton et al., eds., Transforming Enterprises: The Economic and Social Implications of Information Technology. Cambridge, MA: MIT Press, 2005, pp. 49-75. Jorgenson, Dale W.; Landefeld, J. Steven and Nordhaus, William D. A New Architecture for the U.S. National Accounts. Chicago: University of Chicago Press for the NBER, 2006 (forthcoming). Kozicki, Sharon. “The Productivity Growth Slowdown: Diverging Trends in the Manufacturing and Service Sectors.” Federal Reserve Bank of Kansas City Economic Review, First Quarter 1997, pp. 31-46. Kozicki, Sharon. “How Do Data Revisions Affect the Evaluation and Conduct of Monetary Policy?” Federal Reserve Bank of Kansas City Economic Review, First Quarter 2004, 89(1), pp. 5-38. Landefeld, J. Steven and Grimm, Bruce T. “A Note on the Impact of Hedonics and Computers on Real GDP.” Survey of Current Business, December 2000, pp. 17-22. Oliner, Stephen D. and Sichel, Daniel E. “Information Technology and Productivity: Where Are We Now and Where Are We Going?” Federal Reserve Bank of Atlanta Economic Review, Third Quarter 2002, 87(2), pp. 1-13. Orphanides, Athanasios and van Norden, Simon. “The Reliability of Inflation Forecasts Based on Output Gap Estimates in Real Time.” Journal of Money, Credit, and Banking, June 2005, 37(3), pp. 583-601. Pakko, Michael R. “The High-Tech Investment Boom and Economic Growth in the 1990s: Accounting F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W for Quality.” Federal Reserve Bank of St. Louis Review, March/April 2002a, 84(2), pp. 3-18. Pakko, Michael R. “What Happens When the Technology Growth Trend Changes? Transition Dynamics, Capital Growth and the ‘New Economy’.” Review of Economic Dynamics, April 2002b, 5(2), pp. 376-407. Pakko, Michael R. “Changing Technology Trends, Transition Dynamics, and Growth Accounting.” Contributions to Macroeconomics, 2005, 5(1), Article 12. Runkle, David E. “Revisionist History: How Data Revisions Distort Economic Policy Research.” Federal Reserve Bank of Minneapolis Quarterly Review, Fall 1998, 22(4), pp. 3-12. Sherwood, Mark K. “Difficulties in the Measurement of Service Outputs.” Monthly Labor Review, March 1994, 117(3), pp. 11-19. Solow, Robert. “Technical Change and the Aggregate Production Function.” Review of Economics and Statistics, 1957, 39(3), pp. 312-20. Solow, Robert. “Investment and Technological Progress,” in Kenneth Arrow, Samuel Karlin, and Patrick Suppes, eds., Mathematical Methods in the Social Sciences. Stanford, CA: Stanford University Press, 1960, pp. 89-104. Solow, Robert. “After ‘Technical Progress and the Aggregate Production Function’,” in Charles R. Hulten, Edwin R. Dean, and Michael J. Harper, eds., New Developments in Productivity Analysis. Chicago: University of Chicago Press for the NBER, 2001, pp.173-78. Stiroh, Kevin J. “Information Technology and the U.S. Productivity Revival: What Do the Industry Data Say?” American Economic Review, December 2002, 92(5), pp. 1559-76. Triplett, Jack E. and Bosworth, Barry P. “Productivity in the Services Sector,” in Robert M. Stern, ed., Services in the International Economy. Ann Arbor, MI: University of Michigan Press, 2001; www.brookings.edu/views/papers/triplett/ 20000112.htm. M AY / J U N E 2006 201 Anderson and Kliesen Triplett, Jack E. and Bosworth, Barry P. “Productivity Measurement Issues in Services Industries: ‘Baumol’s Disease’ Has Been Cured.” Federal Reserve Bank of New York Economic Policy Review, September 2003, 9(3), pp. 23–33. Triplett, Jack E. and Bosworth, Barry P. Productivity in the U.S. Services Sector. Washington, DC: Brookings Institution, 2004. Triplett, Jack E. and Bosworth, Barry P. “Baumol’s Disease Has Been Cured: IT and Multifactor Productivity Growth in the U.S. Services Industries,” in Dennis Jansen, ed., The New Economy and Beyond: Past, Present and Future. Cheltenham: Edward Elgar, 2006 (forthcoming); www.brookings.edu/es/research/projects/ productivity/workshops/20020517_triplett.pdf. Yuskavage, Robert E. “Improved Estimates of Gross Product by Industry, 1959-94.” Survey of Current Business, August 1996, 76(8), pp. 133-55. Yuskavage, Robert E. and Pho, Yvon H. “Gross Domestic Product by Industry for 1987-2000.” Survey of Current Business, November 2004, pp. 33-53. 202 M AY / J U N E 2006 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W The Learnability Criterion and Monetary Policy James B. Bullard Expectations of the future play a large role in macroeconomics. The rational expectations assumption, which is commonly used in the literature, provides an important benchmark, but may be too strong for some applications. This paper reviews some recent research that has emphasized methods for analyzing models of learning, in which expectations are not initially rational but which may become rational eventually provided certain conditions are met. Many of the applications are in the context of popular models of monetary policy. The goal of the paper is to provide a largely nontechnical survey of some, but not all, of this work and to point out connections to some related research. Federal Reserve Bank of St. Louis Review, May/June 2006, 88(3), pp. 203-17. INTRODUCTION Overview I n a number of recent papers, economists have begun to analyze the stability of rational expectations equilibria under learning in microfounded models of monetary policy. Most of these analyses have been in versions of the New Keynesian macroeconomics, as presented most prominently by Woodford (2003a). The goal of this paper is to provide a brief, largely nontechnical survey of some, but not all, of this work and to point out connections to some related research. Origins Learning has been an issue in macroeconomics since the rational expectations revolution swept the field in the 1970s and 1980s. Rational expectations has long been understood as a modeling device: When studying economic outcomes, we economists should think of them as equilibria only if expectations are consistent with actual outcomes. But, how is it that economic actors could come to possess rational expectations if they do not initially possess detailed knowledge concerning the nature of equilibrium in the economy or economic situation in which they find themselves?1 Several key papers in the 1980s, including Bray (1982), Evans (1985), Lucas (1987), and Marcet and Sargent (1989a,b), explored an idea concerning one resolution of this question. The idea was that, indeed, economic actors cannot be expected to initially know the nature of the equilibrium of the economy in which they operate. Instead, they have a perception of the equilibrium law of motion, and they use available data generated by the economy itself to update their perceived law of motion using recursive algorithms, 1 Some of the tenor of the earlier, feisty debate on this question is conveyed by the following quote from an influential paper by Stephen DeCanio (1979, p. 52, italics in original): “Thus, direct computation of rational expectations by flesh-and-blood agents in an actual market situation is impossible in practice.” James B. Bullard is a vice president and economist at the Federal Reserve Bank of St. Louis. This paper is a revised and extended version of remarks originally prepared for the conference, “Heterogeneous Information and Modelling of Monetary Policy,” held in Helsinki, Finland, October 2-3, 2003. The author thanks the Bank of Finland and the Center for Economic Policy Research for sponsoring this event, and Seppo Honkapohja, Massimo Guidolin, and Michael Owyang for helpful comments. Deborah Roisman provided research assistance. © 2006, The Federal Reserve Bank of St. Louis. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W M AY / J U N E 2006 203 Bullard such as recursive least squares. Should the perceived law of motion come to coincide with the actual law of motion of the economy, a rational expectations equilibrium will have been attained. The economic actors will have “learned” the rational expectations equilibrium. This idea also has an appealing practical interpretation. In an actual macroeconomic environment, expectations of all of the key players are influenced by the expectations of the forecasting community. The forecasting community uses econometric models of the economy, recursively updated. Thus, it is not too far-fetched to think that a dynamic like the one described is powerful and at work in observed macroeconomies. The question of whether such a process will actually converge or not is technically demanding because, in economic models, beliefs concerning the future help determine actual values of key variables; but, under learning, these same values are used in the recursive updating and so feed back into the generation of updated beliefs. It is not at all clear how such a system should be expected to behave. The findings of Marcet and Sargent (1989a,b) on this question were revised, extended, and explored in a series of papers by George Evans and Seppo Honkapohja during the 1990s. Much of that effort is discussed in the landmark book by Evans and Honkapohja (2001), where they present a complete theory of the effects of recursive learning in macroeconomic environments. One theme of their theory is that local convergence in such systems can often be assessed by calculating a certain expectational stability (E-stability) condition, viewing the mapping from the perceived law of motion to the actual law of motion as a differential equation in notional time. They show the conditions under which the stability of this differential equation governs the stability of the system under real-time recursive learning.2 These conditions are generally quite weak, and so many authors now routinely calculate expectational stability conditions as a means of assessing stability under recursive learning in models of interest. 2 The systems under real-time learning are stochastic difference equations with time-varying coefficients. 204 M AY / J U N E 2006 A Minimal Criterion It is important to stress that the idea of stability under recursive learning—learnability—just outlined can be viewed as a “minimal deviation from rational expectations” approach to this question. The agents in the model are endowed with a perceived law of motion which, in most cases, corresponds in form to the equilibrium law of motion for that economy. Thus, the agents are given the correct specification for their recursively estimated vector autoregressions that they use to forecast the future. In addition, the theorems are local in nature, so that we think of the systems as initially quite near the rational expectations equilibrium. And, the agents are passive updaters—they simply update the coefficients in their model as new data are produced. Convergence hinges on whether initially small expectational errors are damped or magnified as the economy evolves. One interpretation of this is that the situation is very favorable to allowing the agents to learn the rational expectations equilibrium. If the equilibrium cannot be learned even under these very favorable conditions, then one might be quite pessimistic about the possibility of observing such an equilibrium in an actual economy. Thus, the learnability criterion can be viewed as a minimal stability condition that any reasonable equilibrium should be required to meet. What Has Been Learned So Far? The main messages of the learning literature to date are not difficult to summarize. First and foremost, it is possible in many macroeconomic environments that recursive learning as described above can produce a dynamic that converges to a rational expectations equilibrium. So, some rational expectations equilibria are indeed learnable in this sense. Some initial thinking on this issue suggested that a general case could be made for nonconvergence, and thus that rational expectations equilibrium was not a useful concept. But that argument has been dispelled. A second message, however, is that not all rational expectations equilibria are learnable. Some, in fact, are unstable under the recursive learning dynamic. Furthermore, because this conF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Bullard ception of recursive learning involves a minimal deviation from rational expectations, the unlearnable equilibria are particularly suspect as descriptions of actual economies. One certainly has the impression from much of the economics profession that all rational expectations equilibria are somehow learnable,3 but it turns out not to be true. It is perhaps not hard to imagine now that, for systems like this, the feedback could be too strong and expectational errors could be amplified. The state of affairs is thus that some rational expectations equilibria are learnable while others are not. Furthermore, convergence will in general depend on all the economic parameters of a given system, including the policy parameters (that is, it depends on the entire economic structure). Therefore, an important additional message is that policy can have an impact on whether a targeted rational expectations equilibrium is learnable or not. Policymakers therefore may wish to take into account how a particular policy choice might influence the stability of a targeted equilibrium. This feature of the recent literature has generated considerable interest. One additional message is that there appears to be no clear, general relation between conditions for learnability and conditions for determinacy of rational expectations equilibrium. I will discuss this issue briefly below. Alternative Formulations of Learning In a recent after-dinner speech, eminent economist Charles Goodhart remarked that, in his opinion, most learning in a large macroeconomy comes not from statistical regression of any kind, but from information passed from person to person. Goodhart said, “You ask your uncle.”4 That comment certainly rings true and echoes a longstanding criticism of the learning literature as I have described it. But learning along this line has also been pursued in the macroeconomics and finance literature. A key aspect of the Goodhart comment is that important economic judgment travels from 3 This seems to be the message in Lucas (1987). 4 I am paraphrasing a portion of the remarks by Goodhart (2003). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W person to person, leaving different people in the population with different beliefs most of the time. As an example, consider an individual decision that has important implications for macroeconomics: How much should a household save out of current income, and how should savings be allocated among available assets? It seems undeniable that in actual economies, households obtain information to help them answer these questions by asking those around them and by obtaining professional advice. Households with similar characteristics often have very different savings strategies. This would seem to conflict with most models, in which behavior and expectations are homogeneous. The artificial intelligence literature has produced some models that can address some of these issues.5 The ones that have been investigated in economic contexts are often variants of genetic algorithms. Some prominent examples in the literature include Marimon, McGrattan, and Sargent (1990) and Arifovic (1996). In these models, a standard economic environment is assumed, but agents are allowed to hold initially diverse beliefs concerning a key future variable, such as an expected rate of return on an asset. Agents then make optimal decisions given their expectations, which, aggregated over all of the agents in the economy, produces some actual outcomes—prices and quantities—for the current period. Agent beliefs are then updated using genetic operators. These operators draw on evolutionary principles. First, beliefs that deliver low utility to their owners tend to get replaced with beliefs that deliver higher utility. In addition, agents experiment with alternative beliefs, either ones that are mixes of their own and those of other agents in the economy,6 called crossover, or simply by means of a random change in belief, called mutation. With a new set of beliefs in place, new decisions are made, and new outcomes are produced. The question is then: Will such a process converge to a rational expectations equilibrium of the model? 5 Heterogeneity and learning have been addressed outside the artificial intelligence literature as well. See, for instance, Branch and Evans (2006), Giannitsarou (2003), and Guse (2005). 6 This operator relates to Goodhart’s comment. M AY / J U N E 2006 205 Bullard The papers in the evolutionary learning literature for macroeconomics tend to be computational, as few analytical results are available. The short answer is that, yes, processes like the one I have described can converge to rational expectations equilibria of well-defined models. And again, not all rational expectations equilibria are stable under this type of learning dynamic.7 The genetic algorithm approach departs from the “minimal deviation from rational expectations” ideal of the recursive learning literature and asks the learning dynamic to describe a global search for equilibrium from initial agent behavior that might be nearly random. In this sense, the approach is much more ambitious. It is also more attractive as a model of the type of social learning that seemingly takes place every day in observed macroeconomies. The genetic algorithm approach also puts heavy emphasis on how information diffuses across households in an economy. The nature of the information diffusion is based on the properties of the genetic operators that are assumed.8 Relation to Behavioral Finance Sometimes learning is mentioned in conjunction with the burgeoning behavioral finance literature.9 The behavioral finance approach draws on psychology, especially experiments with human subjects, to document behavior patterns. Take the following case of subjects who undergo observation in psychological studies. They may seem to be persistently pessimistic, for example, during the course of the study. The literature would then seek to postulate these behaviors in models to see whether apparent anomalies in financial data can then be explained.10 The behavioral finance approach, then, is quite different from the learning literature as I have described it. The macroeconomics learning literature asks how rational 7 The Arifovic (1996) paper, for instance, describes a process that does not converge and instead produces endogenously fluctuating exchange rates. 8 For a survey of this literature, see Arifovic (2000). 9 For one summary of work in behavioral finance, see VissingJorgensen (2004). 10 These ideas are not so new; see the volume by Hogarth and Reder (1987). 206 M AY / J U N E 2006 expectations could come about, allowing that agents behave optimally given their expectations. The behavioral finance literature seeks to understand the empirical implications of postulating certain types of seemingly irrational, but laboratory documented, behavior on the part of market participants. A natural question, and one that is sometimes asked, is whether the seemingly irrational behavior can survive over a long period of time or whether instead market participants would learn the rational behavior. Thus, learning is often mentioned in conjunction with behavioral finance, and this seems to be a fruitful area of future research. LEARNABILITY IN MONETARY POLICY MODELS Taylor-Type Policy Rules Consider a small, closed, New Keynesian economy described by Woodford (1999 and 2003a) and Clarida, Galí, and Gertler (1999): (1) zt = Êt zt +1 − σ −1 ⎡⎣ rt − Êt π t +1 ⎤⎦ + rtn , (2) π t = κ zt + β Êt π t +1. These equations are derived from a model in which each infinitely lived member of a continuum of household-firms produces a differentiated good using labor alone, but consumes an aggregate of all goods in the economy. The household-firms price their good under a constraint on the frequency of price change. The first-order conditions for the consumption problem yield equation (1) while those for the pricing problem yield equation (2). The variable πt is the percentage-point time-t deviation of inflation from a fixed target value; zt is the output gap, also in percentage points; rtn is an exogenous shock, usually thought of as being serially correlated; and rt is the deviation of the short-term nominal interest rate from the value consistent with inflation at target and which is under the control of the monetary authority. The parameter β is the common discount factor of the households, σ relates to the elasticity of intertemporal substitution in consumption of the F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Bullard household, and κ relates to the degree of price stickiness in the economy. These parameters are argued to be invariant to contemplated changes in policy. Bullard and Mitra (2002) view the inflation target and the long-run level of output as zero. The notation Êt is meant to indicate a possibly nonrational expectation taken using information available at date t, so that Et without the hat is the normal expectations operator.11 To close the model, one might postulate a simple monetary policy feedback rule of the type discussed by Taylor (1993) and analyzed in the large literature since that paper was published. One could write such a rule as (3) rt = ϕπ π t + ϕ z zt , where ϕπ and ϕz are nonnegative and not both equal to zero. The parameters in the policy rule are particularly interesting as they may have an impact on the nature of the rational expectations equilibrium of the model, and they may also have an impact on the ability of the private sector agents to learn a rational expectations equilibrium. One interesting feature of this model is that expectations enter on the right-hand side of equations (1) and (2). This is a consequence of the microfoundations, in which the household-firms are forward-looking in deciding today’s consumption and today’s prices. This would seem to be an inescapable consequence of the microfounded approach; therefore, we might expect all monetary policy models to have this feature in some form, and thus that the type of analysis discussed below should apply to a wide variety of models of monetary policy and not only to the simple example given here. Bullard and Mitra (2002) studied the model (1)-(3) under both a rational expectations assumption and under a learning assumption using the approach of Evans and Honkapohja (2001). Under rational expectations, a key question is whether rational expectations equilibrium is unique, a.k.a. 11 The microfoundations of the model were developed assuming rational expectations. Preston (2005) has argued that these equations would change under some interpretations of the microfoundations when agents are learning. But Evans, Honkapohja, and Mitra (2003) have argued that, under some reasonable assumptions, these equations would remain unaltered. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W determinate. To calculate determinacy properties, substitute (3) into (1) and write the resulting system in matrix form as y t = α + BÊt y t +1 +ℵrtn , (4) with yt = [zt, πt ]′, α = 0, Ꭽ is a conformable matrix that is not needed in the calculations below, and (5) B= 1 − βϕπ ⎞ ⎛σ 1 . σ + ϕ z + κϕπ ⎜⎝κσ κ + β(σ + ϕ z )⎟⎠ Both zt and πt are free variables in this system, and so both eigenvalues of B need to be inside the unit circle for determinacy to hold.12 Bullard and Mitra (2002) show that the condition for determinacy is (6) ϕπ + (1 − β ) ϕ κ z > 1. This condition is a statement of the Taylor principle, as discussed by Woodford (2001 and 2003a). From equation (2), a permanent increase in inflation increases the output gap by (1 – β )/κ percentage points. Then, given equation (3), the left-hand side of (6) can be interpreted as the extent of the long-run increase in the nominal interest rate in response to a permanent change in inflation. The condition (6) states that this response must be greater than 1, that is, that nominal interest rates must rise more than one-for-one with inflation to achieve determinacy of equilibrium. Even when determinacy obtains, however, the question of learnability still needs to be decided. To calculate learnability, Bullard and Mitra (2002) postulated a perceived law of motion for the private sector given by (7) y t = a + crtn , where a is a 2 × 1 vector and c is a 2 × 2 matrix. This perceived law of motion corresponds in form to the minimal state variable solution to equation (4) and thus endows the agents with the correct specification of the rational expectations equilibrium. Under this perceived law of motion, agent expectations are given by 12 Blanchard and Kahn (1980). M AY / J U N E 2006 207 Bullard Figure 1 Policy Rules with Contemporaneous Data ϕz 4 Indeterminate and E-unstable Determinate and E-stable 3 2 1 0 0 1 2 3 4 5 ϕπ 6 7 8 9 10 NOTE: Regions of determinacy and expectational stability for the class of policy rules using contemporaneous data. Parameters other than ϕπ and ϕz are set at baseline values. Reprinted with permission from Bullard and Mitra (2002). (8) Et y t +1 = a + cρrtn , where ρ is the serial correlation parameter for the shock rtn. Substituting equation (8) into equation (4) yields the actual law of motion given the perceptions in equation (7), namely, (9) y t = Ba + ( Bcρ +ℵ) rtn . Equations (7) and (9), the perception and the reality, respectively, together define a map, T, as (10) T (a, c ) = ( Ba, Bcρ +ℵ). Expectational stability is determined by the matrix differential equation (11) d (a,c ) = T (a,c ) − (a,c ). dτ If the differential equation (11) is asymptotically stable at the fixed point (a–,c– ) the system is said to be expectationally stable. A key result in Bullard and Mitra (2002) is to show that the condition for expectational stability in this system is exactly the inequality (6). As has 208 M AY / J U N E 2006 been argued, this condition corresponds exactly to the Taylor principle applied to this system. Thus, the Taylor principle delivers both determinacy and learnability for a standard New Keynesian model.13 It would seem to be good advice to give to policymakers, both from the point of view of uniqueness of equilibrium and from the point of view of achievability of that equilibrium, that they adopt the Taylor principle in selecting a particular policy rule—values for ϕπ and ϕz—in this model. This key result is summarized in Figure 1 from Bullard and Mitra (2002), where parameter values other than those in the policy rule have been set at the calibration values recommended by Woodford (1999). The message of Figure 1 is that, so long as the monetary authority chooses a pair of values, ϕπ and ϕz , that are sufficiently large, or “aggressive,” then the economy will possess an equilibrium that is both unique and learnable. Should the policymaker choose values in such a 13 For some further discussion of the connections between the conditions for determinacy and those for learnability in this model, see Woodford (2003a,b). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Bullard Figure 2 Policy Rules with Lagged Data ϕz 4 Determinate and E-stable Determinate and E-unstable Indeterminate and E-unstable 3 Explosive 2 1 0 0 1 2 3 4 5 ϕπ 6 7 8 9 10 NOTE: Determinacy and learnability for rules responding to lagged data, with parameters other than ϕπ and ϕz set at baseline values. Determinate equilibria may or may not be E-stable. Reprinted with permission from Bullard and Mitra (2002). way that the Taylor principle (6) is violated, then determinacy does not obtain and unexpected outcomes may arise. Among the pairs of ϕπ and ϕz that deliver determinacy and learnability, policymakers can apply other criteria, such as the expected utility of the representative household, to decide on an optimal policy. More information can be gleaned from Figure 1, however. Under rational expectations, once one demonstrates that a determinate equilibrium exists, there is little further to discuss, other than the quantitative nature of the equilibrium itself. Under learning, however, there is more to the story, because even within the determinate and learnable region, the choice of the parameters in the policy rule will influence the speed with which the private sector can learn the rational expectations equilibrium. This issue has been analyzed in Ferrero (2004). Some policy choices may involve learning times that are extremely long, and hence policymakers may wish to think twice about adopting them. Figure 1 would seem to suggest that determinacy and learnability go hand in hand, but this F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W is not the case. Consider the alternative policy rule defined by (12) rt = ϕπ π t −1 + ϕ z zt −1. Here the monetary authority reacts to last period’s values of inflation deviations and the output gap, perhaps because of realistic information lags. As McCallum (1999) has emphasized, central banks do not observe inflation or the output gap in the same quarter in which they must make decisions regarding their short-term nominal interest rate target. Bullard and Mitra (2002) show that this case is more complicated and in fact that the conditions for determinacy and learnability do not align. This result is shown in Figure 2. The conclusion is that determinacy does not imply learnability. The darkest region in the figure indicates a situation where the policy rule generates determinacy, but not learnability. The policy rules that have been considered so far have the monetary authority reacting to current or past developments concerning key economic variables. But one might imagine that central banks are forward-looking, so that they M AY / J U N E 2006 209 Bullard react not to current or past data directly, but to their own forecast of future developments, say, one period in the future. This case can also be analyzed, assuming that both the private sector and the central bank learn in exactly the same way. Bullard and Mitra (2002) calculate determinacy and learnability conditions in this case and find that the two criteria do not coincide when central banks are forward-looking.14 In a closely related paper, Bullard and Mitra (2006) consider the more complicated, but more realistic, situation when the central bank also includes a lagged interest rate in its policy rule, (13) rt = ϕπ π t −1 + ϕ z zt −1 + ϕ r rt −1, with ϕr > 0. They come to the conclusion that policy inertia tends to improve the prospects for both determinacy and learnability. This might provide some part of an explanation as to why empirical estimates of actual central bank behavior put important weight on the lagged value of the short-term nominal interest rate.15 Optimal Policy Rules Svensson (2003) has argued that postulating Taylor-type monetary policy rules, even with open coefficients16 as in Rotemberg and Woodford (1999), is not a satisfactory practice. Instead, the monetary authority should be modeled as having an objective that they wish to accomplish as best they can with the instruments at their disposal and under the constraints imposed on them by the economic environment. Such an approach would imply “a more complex reduced-form reaction function” (Svensson, 2003, p. 14). One could argue with this conception. By specifying a class of linear policy feedback rules, the analysis can isolate conditions for determinacy and learnability for rules within the class—and then calcu14 For a recent analysis of the related issue of constant interest rate forecasts on the part of central banks, see Honkapohja and Mitra (2005). 15 Typical estimates in the literature put the value of ϕr at 0.7 or even 0.9, depending on the country and the time period. 16 That is, without assigning specific numerical values. Rotemberg and Woodford (1999) indeed found optimal policy rules, but within classes of possible rules that look like the ones Taylor discussed, such as (13). 210 M AY / J U N E 2006 late an optimal rule from among the ones that satisfy the determinacy and learnability conditions according to any criterion one wishes to ascribe to the policymaker. By specifying policymaker behavior according to a given objective first, one risks specifying a policy rule that generates indeterminacy, unlearnability, or both. One example of this phenomenon occurs in Evans and Honkapohja (2006 and 2003a,b). They considered the economy described by equations (1) and (2) but replaced (3) with an explicit optimization problem for the monetary authorities to solve. This problem can be viewed as policymakers attempting to maximize ∞ (14) Et ∑ β s ⎡⎣π t2+ s + α zt2+ s ⎤⎦ , s =0 where β is the discount factor used by policymakers (assumed to be the same as the discount factor used by the private sector) and the relative weight on output versus inflation is given by α , with α = 0 corresponding to the “strict inflation targeting” case.17 The inflation target is again assumed to be zero. It is well-known that the firstorder conditions for this problem differ depending on whether one assumes a discretionary central bank or one that is able to commit to a superior policy by taking a timeless perspective.18 Under discretion the first-order condition is (15) κπ t + α zt = 0, whereas under commitment it is (16) κπ t + α ( zt − zt −1 ) = 0. Evans and Honkapohja (2006 and 2003a,b) stress that one still needs an interest rate reaction function to implement the policy, and, importantly, there are many such functions that will implement the optimal policy under rational expectations. Do all of these possible reaction functions induce equilibria with the same determinacy and learnability properties? In fact, they 17 Woodford (2003a) has argued that objective (14) approximates the utility of the representative household, in which α takes on a specific value. 18 See Woodford (2003a). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Bullard do not. One might consider the “fundamentalsbased” optimal policy—that is, an interest rate rule that calls for instrument adjustments directly in response to the fundamental shocks.19 One can write down such a rule for either the discretionary or the commitment case. The startling result of Evans and Honkapohja (2003a) is that interest rate reaction functions of this type invariably imply that the equilibrium is unstable in the learning dynamics. Equilibrium is also always indeterminate. Evans and Honkapohja (2003b) label this finding “deeply worrying,” and, indeed, the analysis shows the dangers of proceeding naively from the objective (14) to an implementable policy without considering the effects of that policy on the nature of equilibrium or the stability of the equilibrium in the face of small expectational errors. However, equilibrium can be rendered both determinate and learnable with an alternative interest rate feedback rule, as Evans and Honkapohja (2003a) show. This alternative rule still implements the optimal policy according to the objective (14), but it does so in a way that creates a determinate and learnable equilibrium. The key is to augment the set of variables included on the right-hand side of the feedback rule to include private sector expectations of key variables (the output gap and inflation) as well as the fundamental shocks of the model. This alternative representation of the optimal policy rule is successful in generating determinacy and learnability because it does not assume the private sector has rational expectations, instead allowing the central bank to react to small expectational errors. Of course, for this type of policy rule to be of importance in actual economies, one has to assume that private sector expectations are observable.20 Learning Sunspots With the rational expectations revolution came the idea of sunspot, or nonfundamental, equilibria, in which homogeneous expectations 19 For one discussion, see Clarida, Galí, and Gertler (1999). 20 The Evans and Honkapohja (2006 and 2003a,b) results are sensitive to the specification of the objective function. If one includes interest rate deviations in the objective, E-stability can be achieved without requiring the monetary authority to react to private sector expectations. See Duffy and Xiao (2005). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W are conditioned on purely extrinsic uncertainty, that would not matter for the economy except that agents do condition their expectations on it. This idea has had considerable success as interpretations of many macroeconomic events seem to be consistent with the idea of self-fulfilling expectations. A general finding in the theory literature is that sunspot equilibria exist when equilibrium is indeterminate, so that indeterminacy can imply both the existence of multiple, fundamental equilibria and also the existence of additional, nonfundamental sunspot equilibria. But could agents actually learn such equilibria, in the sense we have described here? To do so, the agents would need to have a perceived law of motion that is consistent with the possibility of a sunspot variable playing an important role. In a classic paper, Woodford (1990) addressed this question and argued that, indeed, a simple recursive learning dynamic might lead agents to coordinate on a sunspot equilibrium. His environment was a version of the overlapping generations model. Honkapohja and Mitra (2004) carry out an analysis of the learnability of nonfundamental equilibria in models like the one described in equations (1)-(3). They find that the Taylor principle continues to play an important role in the learnability of nonfundamental equilibria. In their analysis, violations of the Taylor principle tend to imply indeterminacy, and none of the equilibria are learnable in those cases. Thus, violation of the Taylor principle would seem to imply that the private sector cannot coordinate on a rational expectations equilibrium of any kind in the context of the New Keynesian model.21 This idea turns out not to completely characterize the situation, however. Evans and McGough (2005) show that sunspot equilibria may indeed be learnable if one focuses on common factor representations of the sunspot solution. The tendency in the monetary policy literature, and indeed in the macroeconomics theory literature generally, has been to regard the case of indeterminacy and possible sunspot equilibria as a situation to be avoided at all costs. If a par21 Similar results occur in a real business cycle context with indeterminacy. The sunspot equilibria that exist there are generally not learnable, as shown by Duffy and Xiao (2006). M AY / J U N E 2006 211 Bullard ticular policy generates indeterminacy, then in the eyes of most authors the policy is not a desirable one, quite apart from any question concerning learnability of equilibrium. A dissenter from this view is McCallum (2003), who argues that when multiple equilibria exist, only fundamental, minimal state variable solutions are likely to be observed in practice, and thus arguments based on the mere existence of many nonfundamental equilibria should be given less weight in the literature. A portion of his argument is that nonfundamental equilibria are unlikely to be learnable. In discussing McCallum, Woodford (2003b) argues that, because in the indeterminate cases the minimal state variable solution is also often not learnable, as in the Honkapohja and Mitra (2004) analysis, one should not rely solely on the minimal state variable criterion in generating a “prediction” from a given model. LEARNABILITY IN RELATED MODELS Liquidity Traps The fact that Japan has experienced zero or near-zero short-term nominal interest rates for several years has rekindled ideas about liquidity trap equilibria originally discussed in the 1930s. Benhabib, Schmitt-Grohé, and Uribe (2001) presented an influential analysis of this situation. They argued that the combination of a zero bound on nominal interest rates, commitment of the monetary authority to an active Taylor rule (that is, one that follows the Taylor principle) at a targeted level of inflation, and a Fisher relation generally implies the existence of a second steadystate equilibrium. This second steady state is characterized by low inflation (lower than the target level) and low nominal interest rates in a wide class of monetary policy models currently in use. The Taylor principle does not hold at the lowinflation steady state. They also showed, in the context of a specific economy, the existence of equilibria in which interest rates and inflation are initially in the neighborhood of the targeted inflation rate, but which leave that neighborhood 212 M AY / J U N E 2006 and converge to the low-inflation steady state. From the perspective of the literature on expectational stability, a natural question is, Which of the steady-state equilibria presented by Benhabib, Schmitt-Grohé, and Uribe (2001) are learnable? Based on the results presented so far, in which the Taylor principle governs convergence under recursive learning, one might expect that the targeted, high-inflation equilibrium (in which the Taylor principle holds) would be stable under recursive learning, while the low-inflation equilibrium would not be. Evans and Honkapohja (2005) analyze versions of the Benhabib et al. (2001) economy in which this logic generally holds. The monetary authority in Evans and Honkapohja (2005) can switch to an aggressive money supply rule at low rates of inflation, and this switch can support a third steady state characterized by an even lower inflation rate. This steady state can be learnable in their analysis, and in this sense they find a learnable liquidity trap. But if the monetary authority switches to the money supply rule in support of an inflation rate that is sufficiently high, then the economy is left with only the targeted, relatively high-inflation steady state as a learnable equilibrium. Another analysis of this issue is by Eusepi (2005), who also finds some instances of a learnable liquidity trap in a model with a forecast-based interest rate rule. Eusepi (2005) also provides an analysis of the nonlinear dynamics of this model under learning. As a border of a stable region of the parameter space is approached (say, as a particular policy parameter is increased), an eigenvalue crosses the unit circle, which is normally a defining feature of a local bifurcation. The system can then display cycles and other stationary behavior in a neighborhood of the steady state. Eusepi (2005) finds that this type of outcome can occur in versions of the model studied by Benhabib, Schmitt-Grohé, and Uribe (2001) under learning.22 22 Models with multiple steady states are a natural laboratory for the study of learning issues, independent of questions about liquidity traps. A recent example is Adam, Evans, and Honkapohja (2006). F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Bullard The Role of Escape Dynamics An alternative approach to low nominal interest rate outcomes is studied in Bullard and Cho (2005). Their model is linear and possesses a unique equilibrium in which inflation is near target at all times. To explain persistently, and unintentionally, low nominal interest rates, they design their model to produce an “escape” from the unique equilibrium toward a nonequilibrium focal point, which is characterized by low nominal interest rates and low inflation. The systems they study tend to return to the unique equilibrium following these episodes of “large deviations.” Thus, the Bullard and Cho (2005) approach to low nominal interest rate outcomes does not involve the economy being permanently stuck in a liquidity trap. To generate the escape dynamics, Bullard and Cho (2005) rely on the following features: (i) The private sector has a certain misspecified perceived law of motion for the economy; (ii) there is feedback from the beliefs of the private sector to the actions of the monetary authority; and (iii) the private sector uses a constant gain learning algorithm, which puts more weight on recent observations and less weight on past observations when obtaining key estimates of parameters by means of recursive learning. Students of escape dynamics will recognize the elements just described from themes in Sargent (1999), Cho, Williams, and Sargent (2002), Kasa (2004), Sargent and Williams (2005), and Williams (2001). The escape dynamics in a learning model are interesting because they describe a situation in which the economy is at or near rational expectations equilibrium most of the time, but in which rare events can endogenously push the economy away from the equilibrium toward persistent nonequilibrium outcomes. This may be quite valuable in helping economists understand unusual, but important, macroeconomic events, such as market crashes or depressions. One aspect of this type of analysis is that a rare or unusual event precipitates the escape episode. How rare is this event? In some analyses, it may seem implausible to wait for such a rare event to explain an important macroeconomic outcome. However, McGough (2006) suggests that F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W in models that have escape dynamics, one may not have to wait for the rare precipitating event to occur to observe the escape dynamics. Instead, the escape can be triggered by a shock to the underlying fundamentals of the economy. In a version of Sargent’s (1999) economy, the shock is a plausible shift in the natural rate of unemployment.23 The models with escape dynamics therefore have a certain instability, which might be activated by events other than the precise combination of shocks within the model necessary to generate an escape. Learning and Structural Change It has long been emphasized in economics that for one-time, unanticipated developments, learning makes a great deal of sense and rational expectations is inappropriate. That is, for structural change or other important, one-time shocks, the most appropriate analysis would include transitional learning dynamics as private sector and government officials learn the new equilibrium. The empirical evidence on the existence of structural change in macroeconomic time series is quite strong. For instance, most macroeconomic time series display a reduction in volatility after 1984, according to standard tests. There is a rational expectations approach that one can take to study problems of this kind, such as the one used by Andolfatto and Gomme (2003). One can postulate that a key feature of the economy follows a regime-switching process, with given transition probabilities. One can then compute optimal behavior of the agents in the economy, given that underlying fundamentals may switch between two regimes. A full-information, rational expectations approach would endow the agents with knowledge of the current state along with the probability transition matrix and allow them to make optimal decisions given the uncertainty they face. A more realistic approach, and the one used by Andolfatto and Gomme (2003), asks the agents to infer the regime using available data and knowledge of the transition probabilities. The agents can solve this signal extraction prob23 See Ellison and Yates (2006) for an alternative explanation of the timing of the escape dynamics described by Sargent (1999). M AY / J U N E 2006 213 Bullard lem optimally using Bayesian methods, and this is sometimes thought of as a type of “learning” analysis. However, in the context of the macroeconomic learning literature, this approach is really one of rational expectations given information available to the agents in the model.24 The rational expectations regime-switching approach is interesting, even brilliant, because it transforms an otherwise nonstationary problem into a stationary one, allowing the researcher to maintain a form of the rational expectations assumption. But I do not think this method is the right one for most types of structural change. Most of the shocks we think we observe are one-time permanent events, widely unexpected, such as the productivity slowdown from the 1970s to the 1990s in the United States. The nature of the event is that the current status quo changes permanently, but not to any well-defined alternative status quo. The new reality is learned only after the event has occurred. For this reason, I think subjecting available models to one-time permanent shocks, and allowing the agents in the model to learn the new equilibrium following the shock, is a better model of the nonstationarity we observe in the data. Of course, for recursive learning to tend to lead the economy toward the new equilibrium, the new equilibrium must be expectationally stable, and this expectational stability must extend to a wide enough neighborhood that the permanent shock does not destabilize the economy completely. To implement this type of learning the literature has turned to constant-gain learning, inspired by the discussion in Sargent (1999). Most learning algorithms have today’s perceptions as yesterday’s perceptions plus a linear adjustment that is a function of the forecast error from the previous period. The coefficient multiplying the forecast error would typically be 1/t, to give equal weight to all past forecast errors. But an agent suspicious of structural change may wish to downweight past forecast errors and put more weight on more recent forecast errors. A simple method of doing 24 There has been recent work that draws tighter connections between classical and Bayesian approaches to learning. See, for instance, Evans, Honkapohja, and Williams (2005) and Cogley and Sargent (2005). 214 M AY / J U N E 2006 this is to change the gain from 1/t to a small positive constant. A more sophisticated method is to use a Kalman filter or a nonlinear filter.25 The agent is then able to track changes in the environment without knowing exactly what the nature of those changes may be. Productivity growth may not simply be switching between high and low, but may visit many other regimes, some of which may never have been observed. The tracking idea equips agents with methods of coping in such an environment. It may well be a better model of structural change in the types of problems macroeconomists try to analyze. For examples of economies with structural change and learning dynamics as I have described it, see Bullard and Duffy (2004), Bullard and Eusepi (2005), Lansing (2002), Milani (2005), Orphanides and Williams (2005), and Giannitsarou (2006). RESOURCES ON THE WEB In this paper, I have provided a limited survey of some of the issues and recent results in the macroeconomics learning literature. Much of this literature has provided commentary on monetary policy issues. The learnability criterion is just beginning to be widely used to assess key aspects of policy that have been difficult to address under a pure rational expectations approach. This survey is far from comprehensive. There are many closely related issues that I have not attempted to address here. As of this writing, interested readers can consult the web page maintained by Chryssi Giannitsarou and Eran Guse at Cambridge University, “Adaptive Learning in Macroeconomics,” which provides a more complete bibliography with up-to-date links: www.econ.cam.ac.uk/research/learning/. REFERENCES Adam, Klaus; Evans, George W. and Honkapohja, Seppo. “Are Hyperinflation Paths Learnable?” 25 McCulloch (2005) provides an analysis of the connections between constant-gain algorithms and the Kalman filter. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Bullard Journal of Economic Dynamics and Control, 2006 (forthcoming). Andolfatto, David and Gomme, Paul. “Monetary Policy Regimes and Beliefs.” International Economic Review, February 2003, 44(1), pp. 1-30. Arifovic, Jasmina. “The Behavior of the Exchange Rate in the Genetic Algorithm and Experimental Economies.” Journal of Political Economy, June 1996, 104(3), pp. 510-41. Arifovic, Jasmina. “Evolutionary Algorithms in Macroeconomic Models.” Macroeconomic Dynamics, September 2000, 4(3), pp. 373-414. Benhabib, Jess; Schmitt-Grohé, Stephanie and Uribe, Martin. “The Perils of Taylor Rules.” Journal of Economic Theory, January/February 2001, 96(1-2), pp. 40-69. Blanchard, Olivier J. and Kahn, Charles M. “The Solution of Linear Difference Models under Rational Expectations.” Econometrica, July 1980, 48(5), pp. 1305-11. Branch, William and Evans, George W. “Intrinsic Heterogeneity in Expectation Formation.” Journal of Economic Theory, March 2006, 127(1), pp. 264-95. Bray, Margaret. “Learning, Estimation, and the Stability of Rational Expectations.” Journal of Economic Theory, April 1982, 26(2), pp. 318-39. Bullard, James B. and Cho, In-Koo. “Escapist Policy Rules.” Journal of Economic Dynamics and Control, November 2005, 29(11), pp. 1841-65. Bullard, James B. and Mitra, Kaushik. “Learning, Determinacy, and Monetary Policy Inertia.” Journal of Money, Credit, and Banking, 2006 (forthcoming). Clarida, Richard; Galí, Jordi and Gertler, Mark. “The Science of Monetary Policy: A New Keynesian Perspective.” Journal of Economic Literature, December 1999, 37(4), pp. 1661-707. Cho, In-Koo; Williams, Noah and Sargent, Thomas J. “Escaping Nash Inflation.” Review of Economic Studies, January 2002, 69(1), pp. 1-40. Cogley, Timothy and Sargent, Thomas J. “Anticipated Utility and Rational Expectations as Approximations of Bayesian Decision Making.” Unpublished manuscript, New York University, March 2005. DeCanio, Stephen. “Rational Expectations and Learning from Experience.” Quarterly Journal of Economics, February 1979, 93(1), pp. 47-57. Duffy, John and Xiao, Wei. “The Value of Interest Rate Stabilization Policies When Agents Are Learning.” Working Paper, University of Pittsburgh, November 2005. Duffy, John and Xiao, Wei. “Instability of Sunspot Equilibria in Real Business Cycle Models Under Adaptive Learning.” Journal of Monetary Economics, 2006 (forthcoming). Ellison, Martin and Yates, Tony. “Escaping Volatile Inflation.” Working paper, University of Warwick and the Bank of England, January 2006. Bullard, James B. and Duffy, John. “Learning and Structural Change in Macroeconomic Data.” Working paper 2004-016A, Federal Reserve Bank of St. Louis, 2004. Eusepi, Stefano. “Comparing Forecast-Based and Backward-Looking Taylor Rules: A ‘Global’ Analysis.” Staff Report No. 198, Federal Reserve Bank of New York, January 2005. Bullard, James B. and Eusepi, Stefano. “Did the Great Inflation Occur Despite Policymaker Commitment to a Taylor Rule?” Review of Economic Dynamics, April 2005, 8(2), pp. 324-59. Evans, George W. “Expectational Stability and the Multiple Equilibria Problem in Linear Rational Expectations Models.” Quarterly Journal of Economics, November 1985, 100(4), pp. 1217-33. Bullard, James B. and Mitra, Kaushik. “Learning about Monetary Policy Rules.” Journal of Monetary Economics, September 2002, 49(6), pp. 1105-29. Evans, George W. and Honkapohja, Seppo. Learning and Expectations in Macroeconomics. Princeton, NJ: Princeton University Press, 2001. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W M AY / J U N E 2006 215 Bullard Evans, George W. and Honkapohja, Seppo. “Expectations and the Stability Problem for Optimal Monetary Policies.” Review of Economic Studies, October 2003a, 70(4), pp. 807-24. Evans, George W. and Honkapohja, Seppo. “Adaptive Learning and Monetary Policy Design.” Journal of Money, Credit, and Banking, December 2003b, 35(6, Part 2), pp. 1045-72. Evans, George W. and Honkapohja, Seppo. “Policy Interaction, Expectations and the Liquidity Trap.” Review of Economic Dynamics, April 2005, 8(2), pp. 303-23. Evans, George W. and Honkapohja, Seppo. “Monetary Policy, Expectations, and Commitment.” Scandanavian Journal of Economics, 2006 (forthcoming). Evans, George W.; Honkapohja, Seppo and Mitra, Kaushik. “Notes on Agents’ Behavioral Rules under Adaptive Learning and Recent Studies of Monetary Policy.” Unpublished manuscript, University of Oregon, 2003. Evans, George W.; Honkapohja, Seppo and Williams, Noah. “Generalized Stochastic Gradient Learning.” Working Paper 2005-17, University of Oregon, 2005. Evans, George W. and McGough, Bruce. “Monetary Policy, Indeterminacy and Learning.” Journal of Economic Dynamics and Control, November 2005, 29(11), pp. 1809-40. Ferrero, Giuseppe. “Monetary Policy and the Transition to Rational Expectations.” Paper No. 19, Society for Computational Economics, Computing in Economics and Finance, August 2004. Giannitsarou, Chryssi. “Heterogeneous Learning.” Review of Economic Dynamics, October 2003, 6(4), pp. 885-906. Giannitsarou, Chryssi. “Supply-side Reforms and Learning Dynamics.” Journal of Monetary Economics, March 2006, 53(2), pp. 291-309. Goodhart, Charles A.E. Remarks made at the conference “Expectations, Learning, and Monetary Policy,” sponsored by the Deursche Bundesbank and the 216 M AY / J U N E 2006 Center for Financial Studies, Eltville, Germany, August 30-31, 2003. Guse, Eran A. “Stability Properties for Learning with Heterogeneous Expectations and Multiple Equilibria.” Journal of Economic Dynamics and Control, October 2005, 29(10), pp. 1623-42. Hogarth, Robin M. and Reder, Melvin W., eds. Rational Choice: The Contrast Between Economics and Psychology. Chicago: University of Chicago Press, 1987. Honkapohja, Seppo and Mitra, Kaushik. “Are Nonfundamental Equilibria Learnable in Models of Monetary Policy? “ Journal of Monetary Economics, November 2004, 51(8), pp. 1743-70. Honkapohja, Seppo and Mitra, Kaushik. “Performance of Inflation Targeting Based on Constant Interest Rate Projections.” Journal of Economic Dynamics and Control, November 2005, 29(11), pp. 1867-92. Kasa, Kenneth. “Learning, Large Deviations, and Recurrent Currency Crises.” International Economic Review, February 2004, 45(1), pp. 141-73. Lansing, Kevin. “Learning About a Shift in Trend Output: Implications for Monetary Policy and Inflation.” Working Paper 2000-16, Federal Reserve Bank of San Francisco, July 2002. Lucas, Robert E. Jr. “Adaptive Behavior and Economic Theory,” in Robin M. Hogarth and Melvin W. Reder, eds., Rational Choice: The Contrast Between Economics and Psychology. Chicago: University of Chicago Press, 1987, pp. 217-42. Marcet, Albert and Sargent, Thomas J. “Convergence of Least-Squares Learning Mechanisms in Self-referential Linear Stochastic Models.” Journal of Economic Theory, August 1989a, 48(2), pp. 337-68. Marcet, Albert and Sargent, Thomas J. “Convergence of Least-Squares Learning in Environments with Hidden State Variables and Private Information.” Journal of Political Economy, December 1989b, 97(6), pp. 1306-22. Marimon, Ramon; McGrattan, Ellen and Sargent, Thomas J. “Money as a Medium of Exchange in an F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W Bullard Economy with Artificially Intelligent Agents.” Journal of Economic Dynamics and Control, May 1990, 14(2), pp. 329-73. McCallum, Bennett T. “Issues in the Design of Monetary Policy Rules,” in John B. Taylor and Michael Woodford, eds., Handbook of Macroeconomics. Volume 1C. New York: Elsevier, 1999, pp. 1483-530. McCallum, Bennett T. “Multiple-Solution Indeterminacies in Monetary Policy Analysis.” Journal of Monetary Economics, July 2003, 50(5), pp. 1153-75. McCulloch, J. Huston. “The Kalman Foundations of Adaptive Least Squares, with Applications to U.S. Inflation.” Unpublished manuscript, Ohio State University, August 2005. McGough, Bruce. “Shocking Escapes.” Economic Journal, 2006 (forthcoming). Milani, Fabio. “Learning, Monetary Policy Rules, and Macroeconomic Stability.” Unpublished manuscript, University of California, Irvine, July 2005. Orphanides, Athanasios and Williams, John C. “The Decline of Activist Stabilization Policy: Natural Rate Misperceptions, Learning, and Expectations.” Journal of Economic Dynamics and Control, November 2005, 29(11), pp. 1927-50. Preston, Bruce. “Learning about Monetary Policy Rules When Long-Horizon Expectations Matter.” International Journal of Central Banking, September 2005, 1(2), pp. 81-126. Rotemberg, Julio J. and Woodford, Michael. “Interest Rate Rules in an Estimated Sticky Price Model,” in John Taylor, ed., Monetary Policy Rules. NBER Conference Report series. Chicago: University of Chicago Press, 1999, pp. 57-119. Sargent, Thomas J. and Williams, Noah. “Impacts of Priors on Convergence and Escapes from Nash Inflation.” Review of Economic Dynamics, April 2005, 8(2), pp. 360-91. Svensson, Lars E.O. “Monetary Policy and Learning.” Federal Reserve Bank of Atlanta Economic Review, Third Quarter 2003, 88(3), pp. 11-16. Taylor, John B. “Discretion versus Policy Rules in Practice.” Carnegie-Rochester Conference Series on Public Policy, December 1993, 39(0), pp. 195-214. Vissing-Jorgensen, Annette. “Perspectives on Behavioral Finance: Does ‘Irrationality’ Disappear with Wealth? Evidence from Expectations and Actions,” in Mark Gertler and Kenneth Rogoff, eds., NBER Macroeconomics Annual 2003. Cambridge, MA: MIT Press, 2004. Williams, Noah. “Escape Dynamics in Learning Models.” Ph.D. Dissertation, University of Chicago, 2001. Woodford, Michael. “Learning to Believe in Sunspots.” Econometrica, March 1990, 58(2), pp. 277-307. Woodford, Michael. “Optimal Monetary Policy Inertia.” Working Paper 7261, National Bureau of Economic Research, 1999. Woodford, Michael. “The Taylor Rule and Optimal Monetary Policy.” American Economic Review, May 2001, 91(2), pp. 232-37. Woodford, Michael. Interest and Prices: Foundations of a Theory of Monetary Policy. Princeton, NJ: Princeton University Press, 2003a. Woodford, Michael. “Multiple-Solution Indeterminacies in Monetary Policy Analysis: Comment.” Journal of Monetary Economics, July 2003b, 50(5), pp. 1177-88. Sargent, Thomas J. The Conquest of American Inflation. Princeton, NJ: Princeton University Press, 1999. F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W M AY / J U N E 2006 217 218 M AY / J U N E 2006 F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W