The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Working Paper Series Learning About Informational Rigidities from Sectoral Data and Diffusion Indices WP 10-09 This paper can be downloaded without charge from: http://www.richmondfed.org/publications/ Pierre-Daniel Sarte Federal Reserve Bank of Richmond Learning About Informational Rigidities from Sectoral Data and Di¤usion Indices Pierre-Daniel Sarte Federal Reserve Bank of Richmond May 2010 Working Paper No. 10-09 Abstract This paper uses sectoral data to study survey-based di¤usion indices designed to capture changes in the business cycle in real time. The empirical framework recognizes that when answering survey questions regarding their …rm’s output, respondents potentially rely on infrequently updated information. The analysis then suggests that their answers re‡ect considerable information lags, on the order of 16 months on average. Moreover, because information stickiness leads respondents to …lter out noisy output ‡uctuations when answering surveys, it helps explain why di¤usion indices successfully track business cycles and their consequent widespread use. Conversely, the analysis shows that in a world populated by fully informed identical …rms, as in the standard RBC framework for example, di¤usion indices would instead be degenerate. Finally, the data suggest that information regarding changes in aggregate output tends to be sectorally concentrated. The paper, therefore, is able to o¤er basic guidelines for the design of surveys used to construct di¤usion indices. JEL classi…cation: E32, C42, C43 Keywords: Information Stickiness, Di¤usion Indices, Approximate Factor Model I want to thank, without implicating them in any way, Adrian Pagan, Ricardo Reis, and Mark Watson for helpful discussions and comments. I am also indebted to Nadezhda Malysheva for outstanding research assistance. The views expressed in this paper are those of the author and do not necessarily re‡ect those of the Federal Reserve Bank of Richmond or the Federal Reserve System. All errors are my own. 1 1 Introduction Data provided by statistical agencies regarding the state of the economy typically lag current conditions. For example, manufacturing data are released with a one month lag by the Federal Reserve Board, revised up to three months after their initial release, and further subject to an annual revision. At the monthly frequency, this data is also quite noisy in a way that partially masks underlying business cycle conditions. Thus, in an alternative attempt to track the business cycle in real time, and to con…rm initial Board data releases, information is also compiled by many institutions and government agencies from qualitative data. The Institute for Supply Management (ISM), for example, constructs a widely used monthly di¤usion index of manufacturing production, based on nationwide surveys, that will be the focus of the analysis. In addition, several Federal Reserve Banks including Atlanta, Dallas, Kansas City, New York, Philadelphia, and Richmond, produce similar indices that are meant to capture real time changes in economic activity at a more regional level.1 A central issue pertaining to these surveys is that gathering information on a large number of sectors in a timely fashion is costly and, given time and resource constraints, the scope of the questions must necessarily be limited. Thus, di¤usion indices constructed by the ISM and Federal Reserve Banks rely on simple trichotomous classi…cations whereby respondents are asked whether a variable, say production for that respondent’s …rm, is “up,”“the same,” or “down” relative to the previous period. The number of respondents can vary over time and the respondents themselves need not be the same from survey to survey. Individual responses are aggregated into proportions of respondents reporting a rise, no change, or a fall in output. Di¤usion indices are then constructed by further converting these proportions into aggregate time series meant to track economic activity. The methods typically used in performing these conversions are discussed in detail in Section 2. While various properties of di¤usion indices, sometimes also referred to as balance statistics, have been studied in some detail, this work has been limited because …rm-level data underlying individual survey responses are either not systematically recorded or not publicly available. It has proven challenging, therefore, to say much about the nature of survey responses, and whether they re‡ect informational rigidities. It has also been di¢ cult to explain why converting qualitative answers in the way suggested by di¤usion indices has proven useful in following economic activity in real time.2 1 Other examples of popular national di¤usion indices include the Employment Di¤usion index constructed by the Bureau of Labor Statistics, and the Di¤usion Index for Industrial Production constructed by the Board of Governors. 2 Di¤usion indices, however, have been used to investigate the extent to which expectations can be considered rational as well as to help forecast economic activity. See Pesaran and Weale (2006) for a comprehensive 2 In this paper, I use sectoral manufacturing data to construct an empirical framework composed of hypothetical survey respondents. Each respondent acts as a spokesperson for a …rm whose output re‡ects both aggregate conditions and conditions speci…c to the sector in which it operates. Methods used to construct di¤usion indices are then applied to these hypothetical respondents to create a synthetic di¤usion index of manufacturing production that can be directly compared to that published by the Institute for Supply Management. The analysis makes two key assumptions. First, information is costly to acquire so that survey respondents are not necessarily aware of their …rm’s exact output at each date. Speci…cally, I allow respondents to update their information set infrequently in the manner suggested by Mankiw and Reis (2002, 2006). Second, as noted in Pesaran and Weale (2006), respondents recognize that some changes in their …rm’s output are not necessarily meaningful so that increases or decreases are reported only when exceeding given thresholds. Under the maintained assumptions, a primary objective of the analysis is to provide estimates of i) the degree of information stickiness, and ii) the thresholds that de…ne perceptions of rises and falls in output, that best describe the ISM manufacturing production index. Using data on manufacturing production in 124 sectors over the period 1972-2009, I estimate that survey respondents update their expectations on average every 16 months. Furthermore, the data also suggest that the degree of information stickiness has fallen over time. Thus, using the onset of the Great Moderation to split the sample, the empirical work estimates an average information lag of 20 months prior to the Great Moderation compared to 13 months during the latter period. These …ndings, therefore, are consistent with a fall in the cost of acquiring information over time. For comparison, previous studies relying on nontruncated surveys and aggregate data have found average information stickiness of roughly 12 months, as presented in Carroll (2003) based on data from the Michigan Survey of Consumers and the Survey of Professional Forecasters (SPF), and 12:5 months, as calculated in Mankiw, Reis and Wolfers (2003) using similar information as well as the Livingston Survey. In a di¤erent vein, Mankiw and Reis (2006), and Reis (2009), use quantitative models to calibrate the degree of information stickiness by targeting di¤erent aspects of the cyclicality of aggregate variables or their responses to shocks. This work …nds that information lags of 6 to 16 months are generally most consistent with salient features of the data. Most recently, Coibion and Gorodnichenko (2009) present evidence against the hypothesis of full information based on a variety of survey forecasts. Their …ndings suggest that forecast errors are persistent with a half-life of up to 16 months. More important, the dynamics of these forecast errors are consistent with predictions of models with informational rigidities. None treatment of survey expectations. See also Ivaldi (1992), as well as Jeong and Maddala (1996), for studies of the rationality of survey data: 3 of these studies, however, include the detailed level of disaggregation exploited here. A key implication of this paper is that informational rigidities provide a foundation for the widespread use of di¤usion indices as contemporaneous economic indicators. In particular, these rigidities mean that a considerable fraction of respondents answer surveys based on what they expect their …rm’s output to be given their most recent information rather than actual production. Therefore, high frequency output ‡uctuations that are unrelated to business cycles tend to be …ltered out. Accordingly, around 60 percent of the variation in the monthly ISM production di¤usion index is located at business cycle frequencies compared to just 23 percent of the variance in monthly aggregate manufacturing. Information stickiness, therefore, in e¤ect lets respondents abstract from “noisy”movements in sectoral production. The analysis further shows that in a world populated by identical …rms that are always fully informed, as in the standard Real Business Cycle (RBC) environment for example, di¤usion indices become degenerate and thus cease to be useful.3 As mentioned earlier, information on individual survey responses that underlie the construction of di¤usion indices is either not recorded systematically or not made publicly available. In contrast, because the empirical framework involves the modeling of hypothetical survey respondents, it allows for the tracking, as a counterfactual of sort, of the proportions of “up,”“same,”and “down”responses over past business cycles. The empirical work then shows some notable di¤erences in the historical behavior of these proportions. Prior to the Great Moderation, the proportion of “optimists” (those reporting expected output increases) and “pessimists” (those reporting expected output decreases) play an equal role in driving the di¤usion index. Hence, recessions and expansions are re‡ected by corresponding spikes in the measure of “optimists” and “pessimists.’ However, after the onset of the Great Moderation, while recessions are still marked by sharp increases in the proportion of “pessimists,” movements in the proportion of “optimists” are considerably more subdued. Therefore, from this standpoint, the Great Moderation period is not only associated with a noticeable decline in volatility but also with an asymmetry in sectoral output ‡uctuations. Finally, drawing on previous work in Foerster, Sarte and Watson (2008), the analysis suggests that information regarding changes in overall manufacturing tends to be concentrated in relatively few sectors. Hence, taking as given the methods by which qualitative survey responses are converted into a quantitative di¤usion index, the empirical framework o¤ers 3 There is also a large literature that examines the pitfalls associated with ignoring the distinction between real time and revised data. These problems motivate in part the interest in creating di¤usion indices. See Croushore (2009) for a recent and comprehensive survey of real-time data analysis. See also Runkle (1998), Croushore and Stark (2001), and Fernald and Wang (2005), for the challenges posed by data revisions to the making of policy in real time. 4 some basic lessons regarding the design of surveys that underlie these indices. First, contrary to standard practice at some Federal Reserve Banks, it is not necessary for surveys to try to capture a representative sample of all manufacturing. The intuition is straightforward. In some sectors, variations in output are driven almost entirely by aggregate factors while, in other sectors, output movements re‡ect mostly sector-speci…c considerations. Therefore, to gain insight into current aggregate business cycle conditions, surveys should emphasize the former sectors and largely disregard the latter sectors. Second, having identi…ed sectors whose variations re‡ect mostly factors driving aggregate changes, I show that a useful di¤usion index may be produced using as few as 15 sectors instead of all 124 sectors in the data set. The rest of this paper is organized as follows. Section 2 describes the methods typically used to construct the ISM and other di¤usion indices based on qualitative surveys. Section 3 highlights key di¤erences between sectoral manufacturing production data and the ISM manufacturing production index. Section 4 then presents an empirical framework aimed at reconciling these di¤erences under the assumption that survey respondents update their expectations only infrequently. The estimation methods and …ndings are reviewed in section 5. Section 6 performs a series of counterfactuals that illustrate how the usefulness of di¤usion indices relates to di¤erent aspects of the economic environment. Section 7 o¤ers concluding remarks. 2 Description of the ISM and other Production Di¤usion Indices The Institute for Supply Management is a large U.S. trade association comprising approximately 40,000 supply management professionals. As part of a broader mandate, it compiles a monthly Manufacturing Report on Business based on questions asked of purchasing executives. To keep the survey process straightforward, and to limit the burden on respondents, questions are posed in a format such that they reply with only one of three answers to indicate a change relative to the previous month. The spirit of the survey, therefore, is very much to capture some notion of changes in output otherwise re‡ected more formally in growth rates. In this case, answers regarding production are limited to “up,” “the same,” or “down,”and an index is then constructed from the responses. Because this simpli…cation lets respondents answer more quickly than if a precise answer regarding production changes (rather than a general assessment) were required, it is crucial to the timeliness of the index. The ISM calculates its index by adding the percentage of positive responses to half of the 5 percentage of “same”responses. Formally, let M represent the number of manufacturing sectors that make up total manufacturing as classi…ed by the U.S. Census Bureau. Let xij t denote the output of a given …rm i working in a sector j at date t, and xij t denote its growth rate relative to the previous period. Consider a survey that asks a sample of N respondents in each of these M manufacij ij turing sectors whether their …rm’s output is “up” (uij t ), “the same” (st ), or “down” (dt ), relative to the previous period. Following Pesaran and Weale (2006), the ISM surveying process would then typically be described as cataloging respondents’perception of changes in their …rm’s output at t relative to t if ij ij ij xij t > ; respondent i reports “up”; ut ( ) = 1; st ( ) = dt ( ) = 0; xij t if if 1 in the following manner: xij t < ij ij ; respondent i reports “same”; sij t ( ) = 1; ut ( ) = dt ( ) = 0; ; respondent i reports “down”; The interval [ dij t ( ) = 1; uij t ( )= sij t ( (1) ) = 0: ; ] de…nes an indi¤erence region that represents respondents’ latent perceptions of rises and falls in output. It captures the idea that changes in output may not always be substantive enough to convey meaningful information, or that respondents may not be certain that they are, and therefore may not worth reporting as “up” or “down.” An immediate implication is that whether an output change is considered “up,”“same,”or “down,” depends intrinsically on the threshold that de…nes the bounds of the indi¤erence ij ij interval. This dependence is made explicit by writing uij t ( ), st ( ), and dt ( ) in equation (1). Given the structure of the surveys, the fraction of “optimists”in the sample is given by Ut = M 1 N 1 M X N X uij t ( ): (2) j=1 i=1 Similarly, the fractions of “same”respondents and “pessimists”are given by St = M 1 N 1 M X N X sij t ( ) (3) dij t ( ); (4) j=1 i=1 and Dt = M 1 N 1 M X N X j=1 i=1 respectively. The value of the ISM di¤usion index at t, denoted It , is then de…ned as 1 100 Ut + S t 2 M X N X 1 ij 1 1 = M N uij t ( ) + st ( ) 2 j=1 i=1 It = 6 100: (5) The resulting index values range from 0 to 100, with numbers above 50 generally indicating an expansion of economic activity. In the case of the Federal Reserve Banks (FRB) surveys, the respondents are also asked to report only “increases,” “decreases,” and “no change” in output relative to the previous month, but the form of the index varies slightly relative to the ISM index. The FRB Richmond survey, for example, calculates its index by subtracting the percentage of negative responses from the percentage of positive responses, producing the so-called “balance statistic” motivated by the probability approach of Carlson and Parkin (1975). Hence, in this case, we have that It = (Ut Dt ) 100 M X N X 1 1 = M N uij t ( ) dij t ( ) 100; (6) j=1 i=1 which is bounded between 100 and 100 and takes on a value of zero when an equal number of respondents reports increases and decreases. It is useful to note that actual changes in aggregate manufacturing output, denoted xt , are given by xt = where xjt = P ij i wt M X wtj xjt ; (7) j=1 xij t represents output growth in sector j, wtij is the share (or weight) of …rm i’s production in sector j, and wtj is the share of sector j’s output in aggregate production. Foerster, Sarte and Watson (2008) show that movements in xt are relatively invariant to the exact sectoral weighting scheme so that the expression in (7) is well approximated P P ij ij P xt . Therefore, if the sample of respondents, N , is xjt = M 1 M by M 1 M j=1 i wt j=1 large enough, the di¤usion indices in (5) and (6) rely on approximately the same aggrega- tion used to arrive at manufacturing output growth. A key di¤erence, of course, is that the variables being aggregated in the di¤usion indices are truncated reports of individual …rm output changes (in the sense of being translated to 0s and 1s) rather than actual output growth. Some key questions that the analysis will address are: i) How well does the ISM di¤usion index of manufacturing production capture variations at business cycle frequencies and, moreover, how does it compare to actual manufacturing output growth? ii) How is the di¤usion index’s ability to track movements at business cycle frequencies related to various features of the environment, in particular the degree of information stickiness characterizing survey respondents? iii) Given an upper bound on the number of sectors that can feasibly be surveyed in a given period, how does one choose which sectors to survey? Put another way, 7 how does one distinguish between sectors that are informative about the state of aggregate manufacturing and those that are not? 3 Basic Properties of Sectoral Manufacturing Data and the ISM Di¤usion Index Because Federal Reserve Banks’di¤usion indices re‡ect regional rather than national conditions, and given that manufacturing data is unavailable at the state level, the analysis uses nation-wide sectoral manufacturing data and the corresponding ISM manufacturing production di¤usion index. As explained above, the di¤usion index is a monthly series obtained from the Institute for Supply Management constructed as in equation (5). Monthly data on manufacturing production are obtained from the Board of Governors over the period 19722009. The manufacturing sector is disaggregated into 124 industries according to the North American Industry Classi…cation System (NAICS), which corresponds roughly to a six-digit level of disaggregation. The raw output data are used to compute sectoral growth rates as well as the relative shares of di¤erent sectors in aggregate manufacturing. Monthly growth rates (in percentage points) in sectoral output are computed as xjt = ln(Xtj =Xtj 1 ) 1200, where Xtj denotes production in the j th sector at date t. The main properties of the data are described in Table A1. Figures 1A and 1B show the behavior of manufacturing production growth, computed according to equation (7), and that of the monthly ISM manufacturing production index over the period 1972-2009. The intervals de…ned by the dashed vertical lines depict recessions dated by the National Bureau of Economic Research (NBER). Looking at Figure 1A, monthly growth rates in manufacturing productions are quite volatile, exceeding 8 percentage points (at an annual rate) over the whole sample period. The fall in volatility associated with the Great Moderation is also evident in Figure 1A; the standard deviation of manufacturing production growth declines essentially by half after 1984. Aside from having a large standard deviation, observe also that the manufacturing production series is relatively “choppy,”with growth in a given month bearing little relationship to growth in the previous or subsequent months. In stark contrast, despite also re‡ecting monthly reported changes, the ISM manufacturing production di¤usion index shown in Figure 1B is much smoother with high frequency ‡uctuations that are much less apparent. At the same time, the ISM series evidently picks up recessions quite well, with the index falling considerably below 50, the neutral threshold in equation (5), in each recession since that of 1973. Given that the ISM manufacturing di¤usion index is meant to capture economic activity in real time, Figure 8 1B makes clear why it is a popular contemporaneous economic indicator.4 To gain additional insight into the two measures of manufacturing production illustrated in Figures 1A and 1B, Figures 2A and 2B show the power spectra of manufacturing output growth and the di¤usion index (up to frequency =2). On the whole, the spectral shapes shown in Figure 2 are typical of growth rate spectra for real macroeconomic variables, as documented for example in King and Watson (1996); the spectra are low at low frequencies, increase to a peak at a cycle length of approximately 50 months, and then decline sharply at high frequencies. King and Watson (1996) refer to this shape as the “typical spectral shape for growth rates” and it is noteworthy that, despite being based on truncated qualitative responses, the spectral shape of the di¤usion index conforms closely to that benchmark. To interpret the shapes shown in Figures 2A and 2B more speci…cally, it is helpful to recall some key concepts of frequency domain analysis. The Spectral Representation Theorem states than any covariance-stationary series, for example xt in this case, can be expressed as a weighted sum of periodic functions of the form cos(!t) and sin(!t), Z Z xt = + (!) cos(!t)d! + (!) sin(!t)d!; 0 (8) 0 where ! denotes a particular frequency and the weights (!) and (!) are random variables with zero means. The variance of xt can then be subsequently decomposed as Z f (!)d!; var( xt ) = 2 (9) 0 where the power spectrum, f (!), gives the extent of frequency !’s contribution to the total variance of the series. Each frequency, !, is in turn associated with cycles of period p = 2 =!. Following King and Watson (1996), business cycle frequencies are de…ned in this paper as those associated with cycles of periods ranging from 24 to 96 months.5 Thus, the dashed vertical lines in Figures 2A and 2B correspond to frequencies, !, ranging from 0:065 = (2 )=96 to 0:26 = (2 )=24. Two observations stand out in Figures 2A and 2B. First, the business cycle interval indeed contains the peak of the spectrum of manufacturing output growth and, remarkably, that of the ISM manufacturing production index as well. More importantly, consistent with Figures 1A and 1B, it is unmistakable that business cycle frequencies explain a much 4 The ISM series, however, is subject to a minor adjustment each year to re‡ect changes in seasonal factors used to construct the index. 5 This de…nition is in turn based on earlier work by NBER researchers using the methods described in Burns and Mitchell (1947). 9 larger fraction of the variance in the di¤usion index than in manufacturing output growth. In particular, compared to the manufacturing di¤usion index, a substantially greater fraction of the variation in manufacturing output growth is located at high frequencies, thus accounting for the “noisy” aspect of output growth relative to the di¤usion index. The power spectra in Figure 2 imply that the business cycle interval contains close to 60 percent of the overall variance in the di¤usion index compared to just 23 percent of the variance in monthly manufacturing output growth. In that sense, month to month, the manufacturing di¤usion index performs considerably better than actual manufacturing output growth in tracking variations at business cycle frequencies. Of course, it is always possible to use quarterly growth rates of manufacturing output, or …lter the series in some other way, to follow its movements at business cycle frequencies. However, the question then is: why does this issue not arise with the di¤usion index which, similarly to output growth, is based on monthly aggregated reports of individual changes in output?6 The next sections will argue that the answer lies not in the truncating and averaging used in equation (5), but follows from having di¤erentially informed survey respondents. While month-to-month variations in manufacturing output growth shown in Figure 1A are large, variations in growth rates at the sectoral level are even more pronounced. This follows from the fact that, in equation (7), some of the sectoral variation “averages out” in aggregation. Figure 3A indeed shows that, at the six-digit level of disaggregation, the standard deviations of sectoral growth rates can easily exceed 100 percent and, on average, are on the order 43 percent compared to a standard deviation of 8:5 percent in aggregate manufacturing growth. Although …rm-level data are not available, the same reasoning suggests that …rm-level variations in output might be even larger. From that standpoint, therefore, it is unclear that surveying individual …rms in the way carried out by the ISM would produce a useful economic indicator. In fact, the ISM production index not only performs well in capturing downturns and upturns in manufacturing generally, but the magnitude of the di¤usion index is also suggestive of the strength in these cyclical swings. Thus, looking at Figure 3B, most index values are clustered between 45 and 55 as expected, but index values of 35 and below are clearly associated with the most signi…cant falls in output growth in Figures 1A and 1B (i.e. the recessions in the 1970s and 1980s as well as the most recent recession). Tables 1 and 2 summarize the main observations made in this section. Table 1 gives the 6 In addition, since monthly manufacturing output is released with a lag and subject to several revisions, the problem of not having the information available for real time analysis persists. This problem is compounded by the fact that, even if an output measure were available in real time, conventional …lters that successfully isolate business cycle frequencies are two-sided. 10 standard deviations of manufacturing output growth and of the ISM index, as well as the fractions of variance explained by business cycle and higher frequencies in the two series. Table 2 shows the autocorrelations in output growth and the di¤usion index, as well as the cross correlations between the two series at di¤erent leads and lags. Observe the distinct di¤erence between the …rst and second row of Table 2. Consistent with the “choppiness”of the manufacturing series shown in Figure 1A, manufacturing output growth in a given month bears little relationship to growth in previous months. In contrast, this is clearly not so for the manufacturing di¤usion index, whose index values in a given month are highly correlated with index values in previous months. In addition, observe also that manufacturing output growth leads the manufacturing di¤usion index in that the correlations between output growth and the di¤usion index are larger for future values, rather than past values, of the index. An objective of the paper will be in part to explain all of these observations. Given the nature of sectoral output growth in manufacturing, the next section sets up an empirical framework that helps explain the key di¤erences between aggregate manufacturing output growth and the manufacturing di¤usion index discussed in Figures 1 through 3 and Tables 1 and 2. The framework exploits the fact that the di¤usion index derives from aggregated reports of monthly manufacturing output changes. Thus, one of its a central assumption is to allow for a distribution of hypothetical respondents with di¤erentially updated information. The paper then explores what degree of information stickiness helps best reconcile the two series. 4 The Empirical Framework Let output growth of a …rm i operating in a sector j evolve according to xij t = xjt + uit ; (10) where Et 1 (uit ) = 0 8i . In other words, changes in output for a …rm working in sector j re‡ect in part changes in that sector’s conditions and in part …rm-level idiosyncratic disturbances that have zero mean. Each …rm is associated with a spokesperson who reports on changes in her …rm’s output. As in Mankiw and Reis (2002), however, I assume that at any given date, it is costly to determine exactly what a …rm’s production changes are, or for the purpose of the surveys, what portion of a …rm’s production changes are actually informative about the current state. The presumption is that information ‡ows from the factory ‡oor, production process, and other relevant sectoral considerations are imperfect and that the …rm representative responding to the surveys is only infrequently apprised of the exact state 11 of output growth. Formally, at each date and in each sector, a fraction 2 (0; 1) of repre- sentatives are able to update their information set. This implies that in each time period, a fraction of spokepersons have current information, a fraction (1 ) of spokepersons 2 have one-period old information, a fraction (1 ) of spokepersons have two-period old information, and so on.7 As discussed above, survey designers ask a sample of N representatives in each of M sectors whether their …rm’s output increased, decreased, or stayed the same at t relative to 1. Because of informational rigidities, respondents’answers cannot always re‡ect their t …rm’s current output growth. Instead, for respondents who do not have current information, answers to the surveys are based on what they expect current output changes to be conditional on their most recent information, Et k ( xij t ), where t k is the date at which they last updated their information set. Because some respondents base their answers on expected output changes, Et k ( xij t ), rather than actual output changes, xij t , a basic element of the empirical framework concerns their perceptions of sectoral output growth, xjt , in equation (10). To this end, I model changes in sectoral output as xjt = j Ft + ejt ; j = 1; :::; M; Ft = (L)Ft 1 + (11) t; where Ft represents a set of latent dynamic factors common to all manufacturing sectors, is a common disturbance such that Et 1 ( t ) = 0, and ejt is a sector-speci…c shock such that Et 1 (ejt ) factor model in (11) can be expressed as j is a factor loading speci…c to sector j, = 0 8j. In vector notation, the dynamic (12) Xt = Ft + et ; where Xt is an M 0 1 vector of sectoral growth rates, ( x1t ; :::; xM t ), matrix of factor loadings, Ft is an r 1 vector of sectoral shocks, M variance-covariance matrix ee . t is an M r 1 vector of manufacturing-wide factors, and et is an 0 (e1t ; :::; eM t ), that are cross-sectionally weakly correlated with The number of time series observations is denoted by T . As discussed in Stock and Watson (2010), the dynamic factor model in (12) has proven a valuable approach to handling, and modeling simultaneously, large data sets where the number of series approaches or exceeds the number of time series observations, as in this paper’s application. Aside from this strict statistical interpretation, however, Foerster, Sarte, and Watson (2008) also show that equation (12) can be derived as the reduced form solution 7 See Reis (2006) for the microfoundations of this approach to modeling information stickiness. 12 to a canonical multisector growth model of the type …rst developed in Long and Plosser (1983), and further studied in Horvath (1998, 2000), Dupor (1999), and Carvalho (2007). Because these models explicitly take into account input-output linkages across sectors, the “uniquenesses,”et , may not satisfy weak cross-sectional dependence. In particular, while Ft in (12) can generally be identi…ed with common shocks to sectoral total factor productivity (TFP), the et ’s re‡ect linear combinations of the underlying structural sector-speci…c shocks. By ignoring the comovement in “uniquenesses,”the factor model (12) can then overstate the degree of comovement in sectoral output that is attributed to common TFP shocks. Using sectoral data on U.S. industrial production and matching input-output tables, Foerster et al. (2008) show that the internal comovement stemming from input-output linkages is relatively small. Therefore, for the remainder of the analysis, I interpret Ft as re‡ecting aggregate sources of variation in sectoral TFP. With the dynamic factor model (12) in hand, it is now possible to create a “synthetic” manufacturing production di¤usion index. The synthetic index is analogous to that discussed in section 2 but makes explicit that not all respondents have up-to-date information when answering surveys. As a simple example, suppose that Ft = Ft 1 + t, < 1. Then, in each sector j, N respondents know their …rm’s current production change exactly, Et ( xij t ) = j Ft + ejt + uit . Furthermore, under the maintained assumptions, (1 xij t = )N respondents last updated their information set in the previous period and, for these respondents, survey answers re‡ect what they expect current output growth to be given that period’s information, Et 1 ( xij t ) = j 2 j Ft 1 . Similarly, (1 )2 N respondents’answers will re‡ect Et 2 ( xij t ) = Ft 2 , and so on. Observe that, except for the respondents who have current information, only …xed sectorspeci…c characteristics and aggregate factors end up playing a role in the construction of the synthetic index. This is because Et k ( xij t ) = so that only the sector-speci…c factor loadings, lags, k j k j Ft k , j = 1; :::; M , and k = 1; 2; ::: , and the factor components and their Ft k , are ultimately relevant. Thus, for the majority of …rms (assuming that is small), variability arising either from …rm-level shocks or from sectoral shocks tends to be …ltered out as Et k (uit ) = 0 and Et k (ejt ) = 0 8i; j and k = 1; 2; ::: . Put another way, as a result of information stickiness, some …rm representatives can only report what they expect output growth to be instead of actual output growth. It follows that for these respondents, month to month shocks a¤ecting changes in …rm output will not be fully re‡ected in the di¤usion index. However, since the goal of di¤usion indices is precisely to capture aggregate business cycles, this implication of infrequent updating turns out to be particularly useful for this purpose. In addition, because answers based on expected output growth re‡ects past 13 information through j k Ft k , information stickiness may help explain not only the smooth nature of the di¤usion index in Figure 1B, but also why manufacturing output growth leads the index in Table 2. Since individual …rm level data is not available, xij t cannot be computed for the fraction of …rms whose respondents have current information. In that case, I assume that xjt = j xij t = Ft + ejt . In other words, currently informed respondents are assumed to represent …rms whose output growth mimics the sector in which they operate. This allows the empirical framework to abstract from individual …rm variability entirely. However, as made clear by Figure 2A, sectoral output remains quite volatile. Therefore, if …rm-level output volatility is considerably more pronounced than sectoral volatility, then the empirical framework only provides a lower bound for the degree of information stickiness. Put another way, more informational rigidity would then be necessary to …lter out high frequency ‡uctuations in output growth in order to obtain the smooth di¤usion index shown in Figure 1B. Analogously to equation (1), the synthetic ISM surveying process described in this section can be characterized as recording, for each sector j, di¤erentially informed perceptions of changes in output according to the following conditions: kj kj kj if Et k ( xij t ) > , then ut ( ) = 1; st ( ) = dt ( ) = 0; k = 0; 1; ::: if if Et k ( Et k ( xij t ) xij t ) < kj kj , then skj t ( ) = 1; ut ( ) = dt ( ) = 0; k = 0; 1; ::: , then dkj t ( ) = 1; ukj t ( )= skj t ( where, at each date t and in each sector j, Et k ( xij t ) = (13) ) = 0; k = 0; 1; :::; j k Ft k for (1 )k N respondents. The proportions of “up,”“down,”and “same”respondents now depend not only on the threshold that de…nes perceptions of rises and falls in output, , but also on the degree of information stickiness, . Given the empirical set-up, the number of “optimists”and “same” respondents in the survey is given by Ut ( ; ) = M 1 M X 1 X (1 )k ukj t ( ) (14) (1 )k skj t ( ); (15) j=1 k=0 and St ( ; ) = M 1 M X 1 X j=1 k=0 respectively. Therefore, similarly to equation (5), the synthetic di¤usion index for manufacturing production, denoted Iet ( ; ), takes the form 14 Iet ( ; ) = 1 100 Ut ( ; ) + St ( ; ) 2 M X 1 X 1 kj (1 )k ukj M 1 t ( ) + st ( ) 2 j=1 k=0 100: (16) Given this synthetic di¤usion index, the natural question is: what degree of information stickiness, , and indi¤erence threshold, , best describe the actual manufacturing production index created by the ISM? Thus, a and min S( ; ) = ; are chosen to satisfy T X t=1 It Iet ( ; ) 2 (17) : Before moving on to the estimation and …ndings, it is worth summarizing the two key elements of the empirical framework set out in this section. First, respondents who do not have current information answer survey questions based on expected output growth, conditional on their most recent information, rather than actual output growth. Hence, since Et k (uit ) = 0 and Et k (ejt ) = 0 8i; j; and k = 1; 2::: , this feature of information stickiness helps …lter out high frequency ‡uctuations that arise through shocks. Second, to the extent that respondents’answers re‡ect past information through j k Ft k , and because equation (17) is a weighted sum of these information lags, one expects the resulting di¤usion index to be smoother than manufacturing output growth. It is also precisely this mechanism that may allow manufacturing output growth to lead the ISM di¤usion index as shown in Table 2. 5 Estimation and Empirical Findings The estimation of the empirical framework described in the previous section proceeds in two steps. The …rst step involves estimation of the dynamic factor model (12). The second step uses the resulting model estimates to construct a synthetic di¤usion index according to equations (13) through (16) and solves equation (17). In the …rst step, the number of factors in (12) are estimated using the Bai and Ng (2002) ICP1 and ICP2 estimators. The factors themselves and the loadings are then estimated by principle component methods. When M and T are large, as they are in this paper’s application, Stock and Watson (2002) show that principle components provide consistent estimates of the factors. In addition, the estimation error in the estimated factors is su¢ ciently small 15 that it can be ignored when carrying out variance decompositions or conducting inference about . In other words, Fbt , can be treated as data in a second-stage regression or subsequent investigation. In the second step, therefore, estimates of the factors obtained in this way are used in the construction of the synthetic di¤usion index, Iet ( ; ), according to the rules given by (13). Equation (17) is then solved for the degree of informational rigidity, , and the indi¤erence threshold, , that best characterize the actual manufacturing di¤usion index. The Bai and Ng (2002) ICP1 and ICP2 estimators yield 2 factors in the full sample (1972-2009), and the …ndings in this section are based on this 2-factor model. However, for robustness, the analysis was also carried out using 1 and 3-factor models with similar results (not shown). Given equation (12), the factor analysis centers on two main results that will help develop intuition for the behavior of the di¤usion index. First, I denote by R2 (F ) the fraction of aggregate manufacturing variability that is explained by common shocks. In particular, letting w denote the M R2 (F ) = w0 FF 0 w= 1 vector of constant mean shares, 2 x, where 2 x xt = w0 Ft + w0 et so that is the variance of aggregate manufacturing output growth. Second, I also highlight the extent to which the common factors explain output growth variability in individual sectors, Rj2 (F ) = j FF j0 = 2 xj , where 2 xj is the variance of sector j’s output growth. The purpose of this last calculation is to show that in some sectors, ‡uctuations in output growth re‡ect in part aggregate factors while, in other sectors, changes in output result mostly from idiosyncratic considerations. This feature of sectoral data will be key in providing guidelines regarding which sectors to survey in the construction of a manufacturing di¤usion index. The factor model implies a volatility of aggregate manufacturing output growth that is nearly identical to that found in the data, 8:47 percent. More important, the common factors explain 85 percent or the bulk of the variability in aggregate manufacturing output growth: Figure 4A further illustrates this point by plotting manufacturing output growth, xt , and the model’s …tted values of the factor component, w0 Ft . Consistent with the factors’dominant role in driving aggregate variability, the two series track each other closely over the full sample period. It immediately follows that, in order to build a di¤usion index that re‡ects aggregate manufacturing output growth, a practical step involves focusing on particular sectors whose output variability is largely driven by the common factors. To help distinguish sectors along this dimension, Figure 4B depicts the distribution of Rj2 (F ) statistics. The …gure shows that, in fact, common factors typically account for a small fraction of the variability in sectoral output growth (the mean and median Rj2 (F ) are 0:17 and 0:13 respectively). Simply put, sector-speci…c shocks tend to drive sectoral variability. 16 However, Figure 4B also shows that this is not the case of all sectors. The factor component explains more than 40 percent of the variations in output growth in approximately 15 sectors, and Rj2 (F ) is as high as 0:65 in this exercise. It is those sectors, therefore, that are likely to be most informative to surveys used to construct a di¤usion index of manufacturing production. Given these …ndings, equation (17) yields estimates of 0:06 for and 3:04 for .8 In other words, respondents update their information set every 16 and a half months on average, and changes in output are reported as “up” or “down” if they exceed 3 percent. Recall that Figure 2 implied a median standard deviation of 31:7 percent for monthly sectoral output growth. Therefore, relative to that benchmark, the indi¤erence interval for which respondents report “no change”appears remarkably narrow, approximately one tenth of the median sectoral standard deviation. In addition, the extent of information stickiness suggested by this experiment is somewhat longer than that found in previous work, mainly with aggregate in‡ation data. For instance, Carroll (2003) uses the Michigan Survey, a quarterly series on households’in‡ation expectations, as well as the Survey of Professional Forecasters over the period 1981 2000, to estimate individuals’degree of information stickiness in forming in‡ation expectations. He …nds that on average, individuals update their expectations once a year. Similarly, Mankiw, Reis and Wolfers (2003) use the Livingston Survey and the Michigan Survey to estimate the rate of information updating that maximizes the correlation between the interquartile range of in‡ation expectations from the survey data with that predicted by the model in Mankiw and Reis (2002). In this exercise, a vector autoregression (VAR) is estimated using monthly aggregate U.S. data to generate forecasts of future annual in‡ation. The authors then …nd that on average, the general public updates their expectations once every 12:5 months. In other work, Mankiw and Reis (2006), as well Mankiw and Reis (2007), estimate that a rate of information updating that generally ranges from 6 to 16 months helps best match key aspects of business cycle ‡uctuations. Finally, Coibion and Gorodnichenko (2009) use various macroeconomic survey forecasts to show that forecast errors are persistent, with a half-life of up to 16 months, and are consistent with predictions of models with informational rigidities. 8 Since the number of time series observations is …nite in practice, the horizon for k in equation (16) must be truncated at some value, kmax . In this case, kmax is set to 35 which can be thought of as an upper bound on information lags. That is, respondents with potentially older information in (13) form expectations ij according to the information set de…ned by kmax , Et k ( xij t ) = Et kmax ( xt ) 8k > kmax . However, note that when = 0:06, only 10 percent of respondents have information lags that excced 35 months. Thus, increasing kmax does not materially a¤ect the …ndings, although this can only be checked to a point since observations are lost as kmax increases. Finally, the number of lags, L, used in equation (11) to model respondents’ expectations is folded into problem (17). Solving this problem gives that L = 2 helps best decribe the ISM di¤usion index in the sense of minimizing the overall sum of squares, S. 17 Tables 3 and 4 describe basic properties of the synthetic di¤usion index estimated from sectoral data. Looking at Table 3, the synthetic index is not quite as volatile as that actually produced by the Institute for Supply Management. However, the proportions of variance of the synthetic index explained by business cycles and higher frequencies almost exactly match those of the ISM di¤usion index. Recall that the monthly sectoral data at the base of the empirical work re‡ect mainly high frequency, or “noisy,” ‡uctuations (Table 2 and Figure 3A). Therefore, information stickiness in essence …lters out these ‡uctuations to produce an index that instead moves mostly with the business cycle. As indicated in Table 4, the autocorrelations of the synthetic di¤usion index at di¤erent lags, (Iet ; Iet k ), closely match those of the actual index created by the ISM, (It ; It k ). Furthermore, because some survey respondents rely on expectations of output changes conditional on information that has not been updated, information stickiness also helps explain why manufacturing output growth leads the ISM di¤usion index. Thus, Table 4 shows that the cross-correlations between manufacturing output growth and the synthetic index at di¤erent leads and lags, ( xt ; Iet+k ), are generally quite close to those between manufacturing output growth and the actual ISM index, ( xt ; It+k ). Figure 5 summarizes these …ndings graphically. Looking at Figure 5A, the synthetic di¤usion index moves relatively closely with the actual ISM index apart from two notable exceptions. First, the synthetic di¤usion index mostly misses the depth of the recessions of the early 1980s. In contrast, the fall in economic activity associated with these recessions is re‡ected in a large decline in the ISM index. Second, the economic expansion that followed the 1991 recession is marked by particularly large values of the ISM index by historical standards, but is more subdued according to the synthetic index. These two key di¤erences between the synthetic and actual ISM indices explain in large part the lower volatility of the synthetic di¤usion index. Comparing Figures 3B and 5B, the distributions of the synthetic and ISM index values are remarkably alike although, as just indicated, the synthetic index does not quite reproduce extreme values of the actual index at either end of the support. Finally, note that the shape of the synthetic di¤usion index’s power spectrum in Figure 5C closely resembles that of the ISM index in Figure 2B. Not surprisingly, therefore, the proportions of variance in the two series that are explained by speci…c frequencies are remarkably close (as in Table 3). Di¤usion indices, in practice, do not systematically record or make public individual survey responses on which they are based. However, as indicated earlier, the proportions of “optimists” (those reporting expected production increases), and “pessimists,” (those reporting expected production decreases) in the empirical framework are simply given by P P1 PM P1 k kj 1 Ut ( ; ) = M 1 M (1 ) u ( ) and D ( ; ) = M )k dkj t t t ( ), j=1 k=0 j=1 k=0 (1 18 respectively. Similarly, the proportion of respondents reporting no change is given by P P1 St ( ; ) = M 1 M )k skj t ( ). Figures 6A and 6B show the model-implied j=1 k=0 (1 behavior of these proportions over past business cycles. Two features are worth highlight- ing. First, at any given time, most respondents typically report no change. Second, the proportions of “optimists” and “pessimists” behave di¤erently before and after the Great Moderation. Thus, prior to 1984, recessions are marked by spikes in the proportion of “pessimists” while expansions are marked by similar spikes in the proportion of “optimists.” However, after the onset of the Great Moderation, while the recessions of 1991, 2001, as well as the current recession, are still marked by sharp increases in the fraction of respondents reporting expected output declines, increases in the proportion of respondents reporting an expected rise in output is more subdued. From this standpoint, therefore, the Great Moderation period is not only associated with a sharp decline in volatility (Figure 1A) but also an asymmetry in sectoral output ‡uctuations. A conjecture that would be interesting to explore is whether this reduction in “optimism”post 1984 may be indirectly to the slow employment recoveries that followed the 1991 and 2001 recessions, or other notable changes in business cycles. 6 Deconstructing the ISM Di¤usion Index Having described the ISM index and its implications for the degree of information stickiness characterizing survey respondents, this section further deconstructs the index according to various components of the empirical framework. In particular, it asks four questions related to the construction of di¤usion indices: i) how does the extent of informational rigidity among survey respondents a¤ect the behavior of the ISM di¤usion index? ii) how important is the degree of heterogeneity across sectors in producing a qualitative index that helps track the business cycle? iii) is it possible to more e¢ ciently construct a di¤usion index by steering the underlying survey’s e¤orts towards key relevant sectors? iv) Does the Great Moderation have implications for potential changes in the types of sectors that are most informative about aggregate manufacturing and the degree of informational rigidity inferred from the ISM index? 6.1 Fully Informed Survey Respondents Figure 7 shows how changes in the degree of informational rigidity, , and the indi¤erence threshold, , a¤ect the behavior of the synthetic di¤usion index relative to that produced by the ISM. Looking at Figure 7A, assuming that 19 is as estimated in section 5 but that = 0, the synthetic di¤usion index becomes considerably more volatile than that estimated in the previous section. Its standard deviation is now 16:8, more than twice as volatile as that of the actual ISM. This …nding re‡ects the fact that all changes in output, no matter how immaterial, are now always reported as “up”or “down.”Moreover, in this case, the synthetic index with = 0 is considerably less correlated with manufacturing production growth, with a correlation of 0:25, compared to the actual correlation of 0:45 which the synthetic index was able to match in section 5. More interesting is the e¤ect of relaxing the information stickiness assumption. Figure 7B illustrates the estimated synthetic index that obtains when respondents are always fully informed, if = 1. In that case, equations (13) and (16) become xjt > , then ujt ( ) = 1; sjt ( ) = djt ( ) = 0 xjt if if is set as in section 5 but all xjt < , then sjt ( ) = 1; ujt ( ) = djt ( ) = 0; , then djt ( and Iet (1; ) = M 1 ) = 1; M X j=1 ujt ( )= sjt ( 1 ujt ( ) + sjt ( ) 2 (18) ) = 0; 100; (19) respectively. When all respondents are fully informed, the empirical framework is one where answers of “up,” ujt ( ), and “same,” sjt ( ), are independent of information lags, k. In essence, the synthetic di¤usion index (19) now re‡ects contemporaneous sectoral output changes up to the truncation rules described by equation (18). One e¤ect of these truncation rules is to transform what would be an overall measure of manufacturing output growth, P xjt , into an index bounded between 0 and 100. M 1 M j=1 Figure 7B depicts the di¤usion index that obtains under this scenario. Two observations are worth noting. First, the synthetic di¤usion index is considerably more volatile than that produced by the ISM. More striking, the time series properties of this synthetic index now more closely match those of manufacturing output growth instead of the ISM index. The correlation between the synthetic index and changes in manufacturing production is 0:85 instead of 0:45. Furthermore, as indicated in Table 5, the proportions of variance of Iet (1; ) that are attributable to business cycles (as well as shorter frequency ‡uctuations) are essentially identical to those of manufacturing output growth. Finally, Table 6 shows that the autocorrelation properties of Iet (1; ) are now much closer to those of manufacturing output growth than those of the ISM index (as shown in Table 4). In sum, without information stickiness, movements in the synthetic di¤usion index essentially mimic those of manufactur- ing output growth despite the truncation rules de…ned by (18). This …nding is reminiscent of the work in Kashyap and Gourio (2007) who show that it is not necessary to keep track 20 of exact changes in a series, in their case aggregate investment, to capture some of its most salient features. Given the di¤erences between Figures 1A and 1B, I interpret this …nding as prima facie evidence that survey respondents do not report current actual changes in output but rather some notion of changes that incorporates past information.9 Figure 8 illustrates the role of sectoral heterogeneity in the construction of di¤usion indices. In particular, it asks: is it important for sectors to behave di¤erently in order to construct meaningful di¤usion indices? Thus, Figure 8A re‡ects a scenario where all sectors have the same factor loadings, equal to the mean factor loading, and sectoral shocks are shut P j down. In terms of the notation introduced earlier, j = = M 1 M 8j. 10 Therefore, j=1 expectations of current output changes, Et k ( xij t ), are no longer sector dependent and the only source of di¤erences across respondents resides in information lags, Et k ( xij t ) = k Ft k 8 i; j and k = 1; 2; :::: (20) By and large, Figure 8A shows that heterogeneity in information lags alone goes a long way towards producing a synthetic di¤usion index that is close to the actual ISM index. This is in spite of what essentially amounts to a representative …rm assumption. In that sense, informational heterogeneity appears at least as important as heterogeneity in production. The fraction of the synthetic di¤usion index variance explained by business cycle frequencies is now somewhat lower than that of the ISM index, 0:45 instead of 0:58, but still considerably higher than that of manufacturing output growth, 0:23. In addition, the autocorrelation structure of the synthetic index series shown in Figure 8A resembles more closely that of the ISM index than that of manufacturing output growth (not shown). Suppose now that, in addition to …rms behaving identically across sectors, respondents are always fully apprised of current conditions and always able to report changes as either “up”or “down.”In this frictionless environment with a representative …rm, typical of many RBC models for example, we have that either ujt = 1 and djt = 0 8j or vice versa so that, at any date, everyone simultaneously reports “up” or “down.” Hence, the ISM di¤usion index can now only take on two values, 100 or 100, and thus becomes degenerate. As shown in Figure 8B, without heterogeneity of any kind, the standard di¤usion index described in section 2 ceases to be useful. Given that the nature of production is the same in Figures 8A and 8B, the contrast between the two …gures only serves to underscore the importance of heterogeneous information lags across respondents. 9 In addition, to the extent that …rm level output growth is even more idiosyncratic than sectoral output growth, a di¤usion index that re‡ects real time output changes would be even less useful. 10 Under this assumption, the factor component becomes an exact proxy for aggregate output growth in PM manufacturing, Ft = M 1 j=1 xjt . 21 6.2 Choosing Which Sectors to Survey in Creating Di¤usion Indices Section 5 presented estimates from the factor model such that i) the factor component, w0 Ft , accounted for most of the variation in manufacturing output growth, xt , (recall Figure 4A), and ii) sectors di¤ered in the degree to which they were driven by common factors rather than idiosyncratic considerations, (recall Figure 5B). These two observations suggest that information regarding the state of overall manufacturing is likely to be located in some sectors more than others. In fact, Figure 5B suggests that many sectors in the data set likely contribute very little information to a di¤usion index meant to track overall changes in manufacturing in real time. To explore this notion further, it is useful to rank sectors according to their Rj2 (F ) statistic. Output variations in sectors where Rj2 (F ) is close to zero are almost entirely driven by idiosyncratic considerations while those with higher Rj2 (F ) re‡ect in part the e¤ects of common factors. Figure 9 then plots di¤usion indices constructed as in section 5 (i.e. with = 0:06 and = 3:04) but using only the top and bottom 15 sectors ranked by Rj2 (F ). Remarkably, Figure 9A shows that surveying only 15 sectors where the common factors have the greatest role is enough to produce a di¤usion index that is nearly identical to that constructed using all sectors in section 5. This …nding re‡ects the fact that information regarding changes in aggregate manufacturing tends to be concentrated in relatively few sectors.11 Conversely, Figure 9B indicates that a di¤usion index constructed using 15 sectors where the common factors are least important would have a noticeably more di¢ cult time tracking expansions and contractions in economic activity. Because the factor component does not dominate variations in sectoral output in this case, the implied di¤usion index is relatively uniform and varies little over time. That said, even in this scenario, information stickiness remains useful in …ltering out high frequency ‡uctuations resulting from idiosyncratic shocks. Therefore, whatever variation is left in the di¤usion index still tends to move with the business cycle. Tables 7 and 8 give the 15 sectors with highest and lowest Rj2 (F ) statistics used to construct the synthetic di¤usion indices in Figures 9A and 9B, as well as their share or weight in overall manufacturing. Interestingly, Table 7 suggests that a fairly large proportion of the most informative sectors involve metal work in one way or another (e.g. Metal Valves, Architectural and Structural Metal Products, Fabricated Metals: Forging and Stamping, Foundries, Metalworking Machinery, Coating and Engraving, etc.). At the opposite ex11 See Foerster, Sarte, and Watson (2008) for alternative calculations that highlight this feature of sectoral data. 22 treme, Table 8 indicates that many of the least useful sectors in the di¤usion index involve food-related industries (e.g. Fluid Milk, Co¤ee and Tea, Animal Food, Seafood Product and Preparation, Wineries and Distilleries, Soft Drinks and Ice, etc.). That said, one should be cautious in interpreting Figure 9A. While it suggests that using only 15 sectors, chosen according to their Rj2 (F ) statistic, is enough to replicate the di¤usion index created using all sectors in section 5, the empirical framework assumes that enough …rms are surveyed in each sector to recover the entire distribution of information lags across respondents. Given time and resource constraints in the surveying process, this is potentially far from the case. Finally, Tables 7 and 8 indicate that sectors that may be most informative in the construction of a di¤usion manufacturing index tend to represent a larger share of manufacturing. However, this relationship is far from tight so that some of the most informative sectors, such Metalworking Machinery or Coating, Engraving, and Allied Activities, have small weights at 0:05 and 0:08, respectively. In fact, virtually all of the least informative sectors in Table 8 have larger weights. This observation points to the pitfall of assuming that sectors that represent a small share of overall manufacturing must necessarily be uninformative about its state. 6.3 The ISM Manufacturing Production Di¤usion Index and The Great Moderation Of course, the sectoral rankings shown in Tables 7 and 8 are not exact and will change somewhat as the estimation is carried out over di¤erent sample periods. One of the most studied aspects of Figure 1 is the break in the volatility of aggregate manufacturing output growth around 1984, and it is natural to ask whether structural changes in U.S. manufacturing over time have led to changes in the way that information is concentrated across sectors. In fact, when the factor model (12) is estimated before and after the Great Moderation, in particular over the periods 1972 1983 and 1984 2009, the sectors that are most and least informative about aggregate manufacturing tend not to change much. Speci…cally, out of the 30 sectors with the highest Rj2 (F ) statistics over the full sample period, 22 of those sectors can be found in the pre Great Moderation period while 23 are found in the post 1984 sample. Similarly, of the bottom 30 least informative sectors, 18 are found in the pre 1984 period while 25 of those sectors are found post Great Moderation. Finally, section 5 pointed out that the rate of information updating estimated in this paper, around 16 months, is somewhat higher than that estimated in other studies using more aggregated survey work, for instance around 12 months in Carroll (2003). Because the latter paper relies on in‡ation and employment expectations measured by the Michigan Survey 23 of Consumers, the sample period in that work re‡ects mostly the post Great Moderation period, in particular 1981 2000 for in‡ation expectations. Now, observe that the ISM manufacturing di¤usion index in Figure 1B does not experience the dramatic decline in volatility that characterizes manufacturing output growth around 1984. It follows that after that date, all else equal, less information stickiness is needed in order to smooth out high frequency output ‡uctuations (since those are less pronounced) and match the di¤usion index in Figure 1B. Consistent with this observation, estimating the empirical model in section 5 over the 1972 1983 period yields = 0:05 compared to = 0:076 over the 1984 2009 period. In other words, the rate of information updating averages approximately 20 months prior to the Great Moderation and falls to around 13 months after 1984. Because the variability of the ISM index remains approximately unchanged throughout the entire sample, this …nding follows almost mechanically from having to smooth larger manufacturing output growth ‡uctuations in the 1970s which are mostly absent starting in the early 1980s (apart from the most recent recession). Therefore, when estimated over separate subsamples, the empirical model suggests, somewhat intuitively, that the cost of acquiring information has fallen over time. 7 Concluding Remarks This paper has used disaggregated manufacturing data to study survey-based di¤usion indices that aim to capture changes in the business cycle in real time. To keep surveys straightforward, and to limit the burden on respondents, these di¤usion indices are generally constructed from questions that require only one of three qualitative answers to indicate changes in a variable relative to the previous month. The empirical framework then recognizes that in answering these survey questions, respondents potentially use infrequently updated information. The analysis suggests that survey answers underlying the ISM manufacturing production di¤usion index re‡ect considerable information lags, on the order of 16 months on average. Furthermore, it underscores that informational rigidities, in essence, lead respondents to …lter out high frequency output ‡uctuations when answering surveys. The resulting index, therefore, is better able to isolate variations at business cycle frequencies. In that sense, informational rigidities provide a foundation for the widespread use of di¤usion indices as a contemporaneous economic indicators. The analysis further shows that in a world populated by fully informed identical …rms, as in the standard RBC environment for instance, di¤usion indices become degenerate. 24 Finally, the empirical work highlights the fact that information regarding changes in aggregate manufacturing output tends to be concentrated in relatively few sectors. Hence, contrary to standard practice, it is not necessary for surveys to try to capture a representative sample of all manufacturing sectors in order to track changes in aggregate activity. The intuition is straightforward. In some sectors, changes in output re‡ect to a signi…cant extent factors that drive aggregate changes while, in other sectors, output variations are mostly explained by idiosyncratic considerations. The analysis then uses factor analytic methods to provide a ranking of the most and least informative sectors in constructing a di¤usion index of manufacturing production. In particular, it shows that a useful index may be produced using as few as 15 sectors instead of all 124 sectors that make up U.S. manufacturing. 25 References [1] Bai, J. and S. Ng (2002), “Determining the number of factors in approximate factor models,”Econometrica, 70, 191-221 [2] Burn, A. and W. Mitchell (1947), “Measuring business cycles,” New York: National Bureau of Economic Research. [3] Carlson, J. and M. Parkin (1975), “In‡ation expectations,”Economica, 42, 123-138. [4] Carroll, C., (2003), “Macroeconomic expectations of households and professional forecasters,”Quarterly Journal of Economics,118, 269-298. [5] Carvalho, V., (2007), “Aggregate ‡uctuations and the network structure of intersectoral trade,”manuscript, University of Chicago. [6] Coibion, O., and Y. Gorodnichenko (2009), “What can survey forecasts tell us about informational rigidities?,”NBER working paper 14586. [7] Croushore, D., (2009), “Frontiers of real-time data analysis,”Journal of Economic Literature, forthcoming. [8] Croushore, D., and T. Stark (2001), “A real-time data set for macroeconomists,”Journal of Econometrics, 105, 111-130. [9] Dupor, W. (1999), “Aggregation and irrelevance in multi-sector models,” Journal of Monetary Economics, 43, 391-409. [10] Fernald, J., and S. Wang (2005), “Shifting data: a challenge for monetary policymakers,” Federal Reserve Bank of San Francisco, 2005-35. [11] Foerster, A., P.-D. Sarte and M. Watson (2008), “Sectoral vs. aggregate shocks: a structural factor analysis of industrial production,”NBER working paper 14389. [12] Horvath, M., (1998), “Cyclicality and sectoral linkages: aggregate ‡uctuations from independent sectoral shocks,”Review of Economic Dynamics, 1, 781-808. [13] Horvath, M., (2000), “Sectoral shocks and aggregate ‡uctuations,”Journal of Monetary Economics, 45, 69-106. [14] Ivaldi, M., (1992), “Survey evidence on the rationality of expectations,” Journal of Applied Econometrics, 7, 225-241. 26 [15] Jeong, J., and G. Maddala (1996), “Testing the rationality of survey data using the weighted double-bootsrapped method of moments,”The Review of Economics and Statistics, 8, 395-403. [16] Kashyap,A., and F. Gourio (2007), “Investment spikes: new facts and a general equilibrium exploration,”Journal of Monetary Economics, 54, 1-22. [17] King, R., and M. Watson (1996), “Money, prices, interest rates and the business cycle,” The Review of Economics and Statistics,”78, 35-53. [18] Long, J., and C. Plosser (1983), “Real business cycles,” Journal of Political Economy, 91, 39-69. [19] Mankiw, N., and R. Reis (2002), “Sticky information versus sticky prices: a proposal to replace the new Keynesian Phillips curve,” Quarterly Journal of Economics, 117, 1295-1328. [20] Mankiw, N., and R. Reis (2006), “Pervasive stickiness,” American Economic Review, 96, 164-169 [21] Mankiw, N., and R. Reis (2007), “Sticky information in general equilibrium,” Journal of the European Economic Association, 5, 603-613. [22] Mankiw, N., R. Reis and J. Wolfers (2003), “Disagreement about in‡ation expectations,”NBER Macroeconomics Annual, 18, 209-247 [23] Reis, R., (2006), “Inattentive producers, ”Review of Economic Studies, 73, 793-821. [24] Pesaran, M., and M. Weale (2006), “Survey Expectations,” in Handbook of Economic Forecasting, Amsterdam: North-Holland, 1, 715-776. [25] Runkle, D. (1998), “Revisionist history: how data revisions distort economic policy research,”Federal Reserve Bank of Minneapolis Quarterly Review, 3-12. [26] Stock, J., and M. Watson (2002), “Forecasting using principle components from a large number of predictors,”Journal of the American Statistical Association, 97, 1167-1179. [27] Stock, J., and M. Watson (2010), “Dynamic Factor Models,” Oxford Handbook of Economic Forecasting, Oxford University Press, forthcoming. 27 Table A1 Summary Statistics of Sectoral Growth Rates by NAICS Industry Classi…cation, 124 Sectors Sector Animal Food Grain and Oilseed Milling Sugar and Confectionery Products Fruit and Vegetable Preserving and Specialty Foods Fluid Milk Creamery Butter Cheese Dry, Condensed, and Evaporated Dairy Products Ice Cream and Frozen Desserts Animal Slaughtering and Meat Processing Ex Poultry Poultry Processing Seafood Product Preparation and Packaging Bakeries and Tortilla Coffee and Tea Other Food Except Coffee and Tea Soft Drinks and Ice Breweries Wineries and Distilleries Tobacco Fiber, Yarn, and Thread Mills Fabric Mills Textile and Fabric Finishing and Fabric Coating Mills Textile Furnishings Mills Other Textile Product Mills Apparel Leather and Allied Products Sawmills and Wood Preservation Veneer and Plywood Engineered Wood Member and Truss Reconstituted Wood Products Millwork Wood Containers and Pallets Manufactured Homes [Mobile Homes] Prefabricated Wood Building and All Other Miscellaneous Wood Products Pulp Mills Paper and Paperboard Mill Paperboard Containers Paper Bags and Coated and Treated Paper Other Converted Paper Products Printing and Related Support Activities Petroleum Refineries Paving, Roofing, and Other Petroleum and Coal Products Organic Chemicals Industrial Gas Synthetic Dyes and Pigments Other Basic Inorganic Chemicals 28 Weight 0.42 0.77 0.55 1.03 0.38 0.01 0.17 0.16 0.11 0.88 0.45 0.14 1.23 0.18 0.98 0.59 0.45 0.27 1.07 0.22 0.67 0.30 0.35 0.20 1.76 0.33 0.43 0.16 0.07 0.09 0.34 0.09 0.13 0.15 0.08 1.58 0.71 0.38 0.37 2.28 1.79 0.34 1.43 0.21 0.15 0.57 St. Dev. 20.34 24.63 42.61 30.45 7.08 59.67 19.99 52.57 36.50 24.45 24.69 64.90 10.99 67.55 18.70 23.06 50.06 85.54 56.86 47.96 19.04 25.15 43.20 21.38 16.98 23.46 59.18 61.58 42.19 47.19 24.25 21.81 57.06 26.65 29.73 25.73 22.66 35.37 25.59 13.66 26.00 27.19 37.32 31.99 77.57 54.49 Min ‐77.95 ‐102.10 ‐164.10 ‐105.55 ‐26.48 ‐267.00 ‐82.90 ‐231.36 ‐138.26 ‐102.78 ‐89.13 ‐183.23 ‐46.35 ‐482.59 ‐70.06 ‐84.53 ‐263.55 ‐341.73 ‐193.63 ‐243.22 ‐78.45 ‐113.25 ‐186.23 ‐92.27 ‐80.86 ‐147.29 ‐360.33 ‐439.26 ‐216.46 ‐207.97 ‐119.46 ‐76.67 ‐217.45 ‐117.69 ‐198.33 ‐93.49 ‐124.69 ‐156.42 ‐104.31 ‐42.62 ‐144.77 ‐158.31 ‐396.50 ‐187.49 ‐297.87 ‐448.42 Max 73.41 97.04 218.06 126.93 21.94 214.20 97.03 193.13 160.08 157.19 109.23 194.39 46.64 256.91 79.24 147.10 161.54 487.06 240.72 173.63 83.48 75.18 126.54 122.51 60.69 72.74 244.06 292.12 140.13 144.84 72.95 105.69 392.50 68.15 126.75 98.47 119.40 141.08 113.58 50.30 177.58 89.93 243.72 151.15 295.62 337.87 Sector Plastics Materials and Resins Synthetic Rubber Artificial and Synthetic Fibers and Filaments Pesticides, Fertilizers, and Other Agricultural Chemicals Pharmaceuticals and Medicines Paints and Coatings Adhesives Soap, Cleaning Compounds, and Toilet Preparation Other Chemical Product and Preparation Plastics Products Tires Rubber Products Ex Tires Pottery, Ceramics, and Plumbing Fixtures Clay Building Materials and Refractories Flat and Brown Glass and Other Glass Manufacturing Glass Container Cement Concrete and Products Lime and Gypsum Products Other Nonmetallic Mineral Products Iron and Steel Products Alumina Refining Primary Aluminum Production Secondary Smelting and Alloying of Aluminum Miscellaneous Aluminum Materials Aluminum Extruded Products Primary Smelting and Refining of Copper Primary Smelting/Refining of Nonferrous Metal [Ex Cu and Al] Copper and Nonferrous Metal Rolling, Drawing, Extruding, and Alloying Foundries Fabricated Metals: Forging and Stamping Fabricated Metals: Cutlery and Handtools Architectural and Structural Metal Products Boiler, Tank, and Shipping Containers Fabricated Metals: Hardware Fabricated Metals: Spring and Wire Products Machine Shops; Turned Products; and Screws, Nuts, and Bolts Coating, Engraving, Heat Treating, and Allied Activities Metal Valves Except Ball and Roller Bearings Ball and Roller Bearings Farm Machinery and Equipment Lawn and Garden Tractor and Home Lawn and Garden Equipment Construction Machinery Mining and Oil and Gas Field Machinery Industrial Machinery Commercial and Service Industry Mach/Other Gen Purpose Mach 29 Weight 0.70 0.10 0.33 0.50 2.63 0.40 0.13 1.41 0.93 2.29 0.43 0.40 0.11 0.15 0.44 0.18 0.19 0.68 0.12 0.37 1.67 0.05 0.14 0.04 0.19 0.09 0.06 0.07 0.35 0.77 0.50 0.34 1.16 0.58 0.29 0.20 1.05 0.41 1.15 0.17 0.39 0.10 0.44 0.29 0.73 2.17 St. Dev. Min 48.01 ‐405.72 65.28 ‐243.86 55.83 ‐326.64 28.60 ‐121.31 14.88 ‐64.87 40.25 ‐217.57 31.46 ‐113.83 25.74 ‐72.81 25.69 ‐99.87 16.00 ‐102.16 72.31 ‐458.82 26.67 ‐177.39 26.62 ‐178.58 40.23 ‐238.86 23.55 ‐92.92 48.81 ‐226.22 53.02 ‐292.85 28.01 ‐103.06 71.19 ‐448.70 31.89 ‐111.52 63.15 ‐311.12 34.00 ‐234.00 24.54 ‐192.40 71.18 ‐217.41 93.97 ‐699.29 95.80 ‐715.36 135.37 ‐1285.13 84.11 ‐405.04 71.67 ‐327.73 24.05 ‐122.17 19.59 ‐89.44 19.44 ‐83.68 13.70 ‐54.41 24.56 ‐79.38 26.47 ‐89.19 20.49 ‐95.10 22.07 ‐79.57 17.86 ‐138.55 13.18 ‐57.14 32.17 ‐190.44 100.40 ‐808.88 70.30 ‐281.26 112.71 ‐696.01 35.46 ‐137.09 26.89 ‐147.88 13.04 ‐55.75 Max 367.90 287.58 215.32 171.54 48.83 147.49 167.52 98.71 88.83 68.43 730.97 132.57 76.15 173.10 94.66 236.17 236.65 84.81 247.56 125.19 232.73 216.92 68.39 566.08 517.89 528.85 743.14 394.53 269.51 72.36 62.80 103.95 43.04 111.19 91.05 57.73 77.58 48.19 38.84 147.70 543.35 249.34 704.18 133.41 73.05 42.24 Sector Ventilation, Heating, Air‐cond & Commercial Refrigeration eq Metalworking Machinery Engine, Turbine, and Power Transmission Equipment Computer and Peripheral Equipment Communications Equipment Audio and Video Equipment Semiconductors and Other Electronic Components Navigational/Measuring/Electromedical/Control Instruments Magnetic and Optical Medi Electric Lighting Equipment Small Electrical Household Appliances Major Electrical Household Appliances Electrical Equipment Batteries Communication and Energy Wires and Cables Other Electrical Equipment Automobiles and Light Duty Motor Vehicles Heavy Duty Trucks Motor Vehicle Bodies Truck Trailers Motor Homes Travel Trailers and Campers Motor Vehicle Parts Aircraft and Parts Guided Missile and Space Vehicles and Propulsion Railroad Rolling Stock Ship and Boat Building Other Transportation Equipment Household and Institutional Furniture and Kitchen Cabinets Office and Other Furniture Medical Equipment and Supplies Other Miscellaneous Manufacturing 30 Weight 0.71 0.84 0.78 1.50 1.54 0.18 2.32 2.34 0.19 0.33 0.15 0.36 0.88 0.16 0.21 0.47 2.28 0.15 0.21 0.08 0.05 0.08 3.04 2.40 0.76 0.23 0.51 0.16 0.86 0.62 1.22 1.36 St. Dev. Min 58.57 ‐161.66 18.21 ‐97.69 36.79 ‐164.35 23.36 ‐51.56 26.44 ‐208.02 143.23 ‐538.96 27.57 ‐159.64 12.95 ‐37.83 41.36 ‐133.44 28.23 ‐157.19 42.18 ‐194.91 70.30 ‐500.84 21.65 ‐62.60 59.40 ‐213.32 27.62 ‐107.21 21.28 ‐82.05 96.68 ‐667.76 187.80 ‐1736.24 64.98 ‐417.67 98.98 ‐627.44 133.18 ‐857.31 96.89 ‐687.25 36.06 ‐196.79 36.34 ‐306.70 36.14 ‐187.51 42.62 ‐161.57 31.09 ‐151.82 50.94 ‐309.80 19.70 ‐81.92 21.73 ‐67.72 11.99 ‐39.71 13.66 ‐54.00 Max 251.32 38.33 153.17 76.46 205.26 782.25 83.97 56.63 161.54 115.86 218.19 428.10 58.28 268.63 99.73 79.85 628.14 1509.01 212.83 550.81 650.42 374.34 191.45 241.50 229.34 151.64 127.32 248.16 65.33 77.00 59.57 49.97 Figure 1. Aggregate Variations in Manufacturing 31 Figure 2. Frequency Decomposition of Manufacturing Variations 32 Figuer 3. Individual Sector Variations and the Distribution of ISM indices 33 Figure 4. Accounting for Manufacturing Variations Using Common Factors 34 Figure 5. Properties of the Synthetic Di¤usion Index 35 Figure 6. Historical Behvavior of the Proportions of “Optimists”and “Pessimists” 36 Figure 7. Removing Informational Rigidities with Heterogenous Sectors 37 Figure 8. Removing Informational Rigidities with Homogenous Sectors 38 Figure 9. Information Concentration and the Construction of Di¤usion Indices 39 Table 1 Volatility of Output Growth and the ISM Di¤usion Index in Manufacturing 1972-2009 Fraction of Variance at Standard Deviation Business Cycle Frequencies 2 years p Fraction of Variance at High Frequencies 8 years p < 2 years Output Growth 8.48 23.39 69.37 Di¤usion Index 6.91 57.91 19.27 Table 2 Autocorrelation and Cross-correlation Structure of Output Growth and the ISM index Autocorrelations (1972-2009) 0 k 1 ( xt ; xt k ) 1.00 0.35 (It ; It k ) 2 032 3 4 5 6 0.27 0.15 0.10 0.10 1.00 0.92 0.85 0.77 0.68 0.60 0.51 Cross-Correlations (1972-2009) -3 k ( xt ; It+k ) -2 -1 0 1 2 3 0.15 0.22 0.31 0.45 0.51 0.50 0.48 Table 3 Volatility of the Manufacturing ISM Di¤usion and Synthetic Di¤usion Indices 1972-2009 Fraction of Variance at Standard Deviation Business Cycle Frequencies 2 years p 8 years Fraction of Variance at High Frequencies p < 2 years Di¤usion Index 6.91 57.91 19.27 Pseudo Di¤usion Index 4.79 54.10 21.66 40 Table 4 Autocorrelation and Cross-correlation Structure of the ISM Di¤usion and Synthetic Di¤usion indices Autocorrelations (1972-2009) 0 k (It ; It k ) (Iet ; Iet k ) k 1 2 3 4 5 6 1.00 0.92 0.85 0.77 0.68 0.60 0.51 1.00 0.95 0.88 0.78 0.67 0.56 0.47 Cross-Correlations (1972-2009) -3 -2 -1 0 1 2 3 ( xt ; It+k ) 0.15 0.22 0.31 0.45 0.51 0.50 0.48 ( xt ; Iet+k ) 0.24 0.25 0.30 0.45 0.52 0.64 0.64 Table 5 Volatility of Manufacturing Output Growth and the Synthetic =1 Di¤usion Index with Fully Informed Respondents, 1972-2009 Fraction of Variance at Standard Deviation Business Cycle Frequencies 2 years p 8 years Fraction of Variance at High Frequencies p < 2 years Output Growth 8.48 23.39 69.37 Pseudo Di¤usion Index 11.60 25.74 63.88 Table 6 Autocorrelations of Manufacturing Output Growth and the Synthetic Di¤usion Index with Fully Informed Respondents, =1 Autocorrelations (1972-2009) k 0 1 2 3 4 5 6 ( xt ; xt k ) 1.00 0.35 032 0.27 0.15 0.10 0.10 (Iet ; Iet k ) 1.00 0.39 0.38 0.41 0.21 0.17 0.21 41 Table 7 Most Informative Sectors Ranked According to Rj2 (F ) Rj2 (F ) Weight Sector 1. Plastic Products 0.65 1.36 2. Household and Institutional Furniture 0.52 1.22 3. Metal Vales Except Balls and Roller Bearings 0.49 0.62 4. Architectural and Structural Metal Products 0.47 0.86 5. Commercial and Service Industry Machinery 0.45 0.17 6. Other Miscellaneous Manufacturing 0.45 0.51 7. Reconstituted Wood Products 0.45 0.23 8. Fabricated Metals: Forging and Stamping 0.45 0.76 9. Foundries 0.43 2.40 10. Fabricated Metals: Spring and Wire 0.43 3.04 11. Sawmills and Wood Preservation 0.42 0.08 12. Metalworking Machinery 0.41 0.05 13. Coating, Engraving, and Allied Activities 0.39 0.08 14. Textile Furnishings Mills 0.37 0.21 15. Other Electrical Equipment 0.37 0.15 42 Table 8 Least Informative Sectors Ranked According to Rj2 (F ) Rj2 (F ) Weight Sector 1. Aircraft and Parts 0.00 0.42 2. Guided Missile and Space Vehicles 0.00 0.77 3. Fluid Milk 0.00 0.55 4. Co¤ee and Tea 0.01 1.03 5. Dry, Condensed, and Evaporated Dairy Products 0.01 0.38 6. Primary Smelting/Re…ning of Nonferrous Metals 0.01 0.01 7. Farm Machinery and Equipment 0.01 0.17 8. Animal Food 0.01 0.16 9. Seafood Product Preparation and Packaging 0.01 0.11 10. Heavy Duty Trucks 0.01 0.88 11. Wineries and Distilleries 0.01 0.45 12. Soft Drinks and Ice 0.02 0.14 13. Copper and Nonferrous Metal Rolling 0.02 1.23 14. Grain and Oilseed Milling 0.02 0.18 15. Mining and Oil and Gas Field Machinery 0.02 0.98 43