Analysis of 35 years of previously unstudied survey data shows how the American public evaluates the health of the macroeconomy. Survey responses are multidimensional, distinct from indexes of "consumer sentiment," and based mostly on genuine perceptions of economic conditions, not media reports of economic statistics. As such, they contain unique information about current and future values of these statistics, particularly consumption growth, a longstanding focus of the literature. Both "intangibles " and macroeconomic fundamentals explain substantial variation in the survey data; the public equates 2 to 5 percentage points of inflation with 1 percentage point of unemployment. (JEL E32, E27, EO1)

I. INTRODUCTION

Traditional business cycle measurement and theory tends to downplay the role of public perceptions. The health of the macroeconomy is evaluated, with a lag, through formal measurement of fundamentals like gross domestic product (GDP) growth, rather than by the opinion of the public. The objective functions employed by models of macro policy, derived from theory or prescribed ad-hoc, are similarly unshaped by public opinion on the attractiveness of various states of the economy.

For a profession that is enamored of consumer sovereignty and cognizant of the value of idiosyncratic, decentralized information, this is a little surprising. It probably reflects, in part, a lack of available data. In the United States, prior research on the macroeconomic perceptions of

the public has almost exclusively employed the widely available Michigan Index of Consumer Sentiment or its counterpart from the Conference Board, the Consumer Confidence Index, to predict future consumption. But these measures are not intended to be, and have not been interpreted as, general assessments of the national macroeconomy. They combine responses to a variety of questions, about personal finances and the macroeconomy, about employment, consumption, and profitability, about current conditions and the change in those conditions. These questions are sufficiently disparate that it is unclear what, exactly, either index measures or how it should be interpreted (see Merkle, Langer, and Sussman 2004). This ambiguity has probably contributed to a decades-long debate over the usefulness of such indices, which has been compounded by divergent findings on their ability to forecast future consumption (see Golinelli and Parigi 2004; Manski 2004).

To surmount these data limitations, 35 years of responses were assembled from three reputable, national U.S. polls that ask about the current state of the national economy and/or whether economic conditions are getting better or worse. All are conceptually and (we show) statistically distinct from indexes of consumer sentiment, and none has been previously studied.

As shown in this article, these poll data enrich our understanding of the business cycle in three ways. They contain information about fundamental macroeconomic variables, which are reported with a lag and (often) subsequently revised. They establish the role of various macroeconomic fundamentals (and of intangibles) in determining the public's satisfaction with the economy, which informs policy and suggests simple summary measures of economic conditions. And they reveal the multifaceted nature of consumers' macroeconomic perceptions, which span four dimensions across a total of nine survey questions. Section II introduces these surveys and shows how they relate to each other and to the Michigan and Conference Board indices. In Section III, we examine how macroeconomic conditions influence the assessments reported therein; their predictive power is examined in Section IV. Section V concludes.

II. POLLING ON THE STATE OF THE ECONOMY

For decades, American news organizations have asked respondents to assess the national macroeconomy; Table 1 lists the questions, time spans, sample periods, and response options for each. (We are not aware of any analogous surveys conducted abroad.) These surveys have developed in three phases.

In the late 1970s and early 1980s, CBS News, (generally) in conjunction with the New York Times (NYT), and ABC News, in conjunction with the Washington Post, occasionally asked a national sample of the U.S. public about the change in macroeconomic conditions: "Do you think the economy (CBS)/nation's economy (ABC) is getting better, getting worse, or staying about the same?" Each of these surveys was, in isolation, too episodic to be of use to researchers.

This changed in the mid-1980s, when both organizations introduced a question about the level, rather than the change, in macroeconomic conditions, and began reporting the responses several times per year. In December 1985, ABC News began asking: "Would you describe the state of the nation's economy these days as excellent, good, not so good, or poor?" CBS News soon followed with a similar, though differently phrased, question: "How would you rate the condition of the national economy these days--Very Good, Fairly Good, Fairly Bad, or Very Bad?" Unlike ABC News, it continued to ask the "better/worse" question as well. Later a third poll, sponsored by USA Today and conducted by Gallup, also asked both types of questions, using somewhat different wording. The USA Today poll was discontinued in October 2008 and the ABC News poll in February 2010.

In the third, modern phase, the quantity, quality, and availability of data have expanded further. In January 2008, Gallup began daily tracking of both "good economy" and "better/worse" questions, interviewing 500 people each day. This continues to be supplemented by the CBS News polls, now conducted almost every month, and by Bloomberg, which picked up the ABC News survey after a year's hiatus. Data from the first two phases are used in this analysis. Initially, the five survey questions in the second phase were analyzed: the "good economy" questions from ABC News, CBS News, and USA Today, along with the "better/worse" questions from CBS News and USA Today. Finding a strong commonality underlying the responses to each question type, we then integrate the responses, using a method described below, and incorporate phase one data to create two latent variables, one for the "good economy" question and one for the "better/worse" question. Further analysis is then conducted using these latent variables. While each survey begins at a different time, listed in Table 1, our data terminate in February 2010 (or, for the USA Today surveys, October 2008). Time is demarcated in months. Each survey contains at least 1,000 respondents in each month it is conducted, and each is conducted in more than half of the months in its sample period (see Table 1).

The first two responses to each "good economy" question are considered "positive." Figure 1A presents the time series of the fraction of each survey's responses that are positive, while Figure IB presents the time series of the fraction of "better" responses to the "better/worse" questions (along with the change in positive responses to the ABC News "good economy" question). The level of positive responses differs across surveys, consistent with the different phrasing of the questions and response options, but the temporal variation appears to be similar, a strong cyclical component punctuated with higher frequency modulations. The large samples in each monthly survey ensure virtually none of this variance (about 0.1%) comes from sampling error.

The Michigan and Conference Board surveys also present indexes of current conditions, also represented in Table 1. (These and a separate expectations index, not represented in the table, average to form that entity's consumer sentiment index.) These questions are distinctly different: neither survey asks explicitly about the overall macroeconomy, but about "business conditions," "available jobs," and durable goods purchases, and most questions concern the respondent or his/her local area, instead of the country as a whole. While these indexes also have a strong cyclical component, the questions themselves do not have the face validity necessary to represent a general assessment of the national macroeconomy. The ABC News, CBS News, and USA Today poll questions do. (1)

Despite this face validity, and microfoundations by which macroeconomic fundamentals generate dynamic responses to opinion surveys (Lux 2009; Easaw and Ghoshray 2010), we recognize that "economists have been deeply skeptical of subjective statements" (Manski 2004, p. 1337, who argues that such skepticism is unwarranted, as does Bewley 2002). Given this skepticism and the aforementioned conflict concerning the consumer sentiment surveys, it is worthwhile to further support these surveys' legitimacy by establishing construct validity (Litwin 1995). This requires a certain logical consistency across surveys: questions about similar concepts should have similar responses (convergent validity), while the responses to questions about distinct concepts should differ (divergent validity).

In our data, convergent construct validity requires that (1) the responses to the "better/worse" questions are related to differences of the "good economy" responses, and (2) the responses to all three "good economy" questions have a strong underlying commonality, despite differences in wording (and similarly for the "better/worse" questions). Divergent construct validity requires, in turn, that (3) these survey responses are distinguishable from the consumer sentiment indexes, which are conceptually different.

A. Levels and Differences

To compare the "good economy" and "better/worse" questions, we begin by regressing the percentage of "better" responses to the CBS/NYT and USA TodaylGaUup "better/worse" questions on leads and lags of the percentage of positive responses to the "good economy" question asked by ABC News--the only one that is reported each and every month. Six months of leads and 12 months of lags were included. A condensed version, using 2-month intervals, is reported in columns 1 and 3 of Table 2.

In these regressions, positive coefficients near the current date offset negative coefficients for the recent past, suggesting the expected differencing interpretation. These differences are essentially backward-looking, with hindsight that extends several months. To pin down the timing more precisely, we found that 2-month combination of the ABC News "good economy" measure that best explains the responses to each "better/worse" question. The optimal pairs are shown in columns 2 and 4 of Table 2, a (mostly) backward difference of 6 months for the USA Today series and 8 months for the longer CBS/NYT series. Treating these specifications as the null, we then tested the alternative that any of the remaining coefficients were nonzero. For neither series could this null be rejected at conventional levels (in the CBS News survey, F = 0.98, p > .10; in the USA Today survey, F- 1.53, > .10). (2)

In three of the four specifications, we cannot reject a strict differencing interpretation that the coefficients sum to zero. This interpretation and the [R.sup.2] values, which range from 0.5 to 0.7, both suggest a reasonable degree of concordance between the level and difference assessments of the macroeconomy. The fit is not so good, however, that the two measures can be considered synonymous. Accordingly, each is examined below.

B. Underlying Commonalities

For each survey question, the response frequencies--the percent of respondents saying the economy is "excellent," and so forth--can be viewed as governed by three terms: a latent variable common to all respondents (L), which can be considered a scalar index of perceived macroeconomic conditions; a random variate (a) that generates cross-sectional variation in individual responses at any given point in time; and a set of thresholds ([mu]) that distinguish an "excellent" response from a "good" response, and so on. If all "good economy" series have a strong underlying commonality, then the latent variables underlying each series should be highly correlated.

Each latent variable, and the associated thresholds, can be estimated by relating the response frequencies nonparametrically to time. The most practical way to do this is to express time as a series of splines, which are used as independent variables in an ordered probit model in which the response frequencies are the dependent variable. (See Takezawa 2006 and Wasserman 2006. Informally, these splines resemble an overlapping sequence of bell curves. We use many splines, 52 over the sample period, to preserve all but the highest frequency variation.) Applying the estimated coefficients to the splines yields a smoothed, unrestricted estimate of the latent variable that extends for the full time span of the survey, filling in any survey-less months.

The formal statement of this model is as follows. For each survey question Z, the individual-level latent variable (/) underlying any discrete choice model equals the sum of L and [alpha], as follows:

(1) [I.sub.j,t] = [L.sup.Z.sub.t] + [[alpha].sub.j,t] = [summation] [[beta].sup.Z.sub.s][S.sub.s,t] + [[alpha].sub.j,t], with [summation over (s)] [S.sub.s,t] = 1 [for all]t and [[alpha].sub.j] ~ N (0,1) [for all]t

Least Favorable Response iff [I.sub.j,t] < [[mu].sup.z.sub.0]

Next more Favorable Response iff

[[mu].sup.z.sub.1] ? [I.sub.j,t] [greater than or equal to] [[mu].sup.z.sub.0]

....

Most Favorable Response iff [I.sub.j,t] [greater than or equal to] [[mu].sup.z.sub.MAX] [[mu].sup.z.sub.0] = 0 [for all] Z

where S is a set of "B-splines," determined according to the method of de Boor (1978), which sum to one at each point in time, and the [mu]s are the thresholds that [I.sub.j,t] must exceed in order for that respondent to report that economic conditions are "excellent" instead of "good," and so on. The predicted value of L at any time T is simply [summation][[??].sub.s][S.sub.s,T].

Figures 2A and 3A present the results for all "good economy" and "better/worse" questions, with the estimated latent variables and thresholds vertically scaled (additively) so that all latent variables have the same mean for the periods in which they overlap. (We do not present confidence intervals, as they are so small.) For both questions the latent variables are as similar as the cutoffs, driven by the various response options, are different. Correlations between the "good economy" latent variables are each 0.99, and the correlation between the "better/worse" latent variables is 0.96.

C. Comparison with Other Indexes

These latent variables are, in effect, indexes themselves, and can be statistically compared to the Michigan and Conference Board indexes. We begin with a simple correlation analysis, presented in Table 3. To incorporate both the "good economy" and "better/worse" series and address potential stationarity concerns, 8-month backward differences are taken of all variables except the two "better/worse" series, consistent with our findings above (for the longer CBS News "better/worse" series). The nine series included in this analysis can be placed into four groups, each clearly demarcated in the table: "good economy" series, Michigan and Conference Board current conditions indexes, expectations indexes, and "better/worse" series. This grouping is supported by the correlation matrix, in which intra-group correlations, of about 0.9, comfortably surpass the cross-group correlations, which never exceed 0.82. (The one exception, the weak correlation between the Michigan and Conference Board current situation indexes, proves the rule. These are composed of responses to quite different questions.) These four groups of variables appear to represent multiple dimensions of macroeconomic perceptions.

To explore further, we employ a principal component analysis. (This has been previously used for quantifying the state of the macroeconomy in other contexts: see the discussion and references in Bai 2003.) This yields a set of nine independent components that linearly reconstruct each of the original series. Economic significance is restricted to the first four of these, which together explain 95% of the aggregate variance of all nine series. The associated factor loadings, also found in Table 3, are easily interpreted. The first, dominant component, a simple, almost-unweighted average of the nine series, reflects basic business cycle variation. The second component distinguishes measures of expectations from everything else. The third component distinguishes the "better/worse" series, while the fourth component represents a difference between the "good economy" series and the Michigan/Conference Board current conditions indexes. The cyclical component explains 72% of the joint variance of these series. Of the remaining 28%, almost half is contributed by the distinctiveness of the expectations indices, and another half by the distinctiveness of the two survey questions analyzed here, with the remainder noise, or "scree."

These findings all support construct validity. The economic assessments that underlie responses to the "good economy" and "better/worse" questions are insensitive to minor differences in wording, logically consistent with each other, and distinct from surveys of consumer sentiment.

III. MACROECONOMIC FACTORS AFFECTING ASSESSMENTS OF THE NATIONAL ECONOMY

A. Aggregated Latent Variables

The results in the previous section indicate that we can adequately describe our survey data with two temporal latent variables, one underlying the responses to all "good economy" questions, the other underlying the responses to all "better/worse" questions. Differences in response frequencies across surveys are captured by differences in the thresholds separating the response options in each survey. This "aggregated" latent variable also improves the temporal coverage for the "better/worse" question, which has occasional temporal gaps in any one survey, which are now "filled in" by another survey. These gaps were particularly frequent prior to 1988, but with this technique, we can incorporate the early CBS News and ABC News responses and extend this latent variable back to 1976, covering a period of high inflation and volatile economic growth.

For each question type, the specification of the model used to estimate this latent variable is a slight modification of that in Equation (1): [L.sub.t.sup.Z] = [L.sub.t] [for all] Z, [[beta].sub.s.sup.Z] = [[beta].sub.s] [for all] Z, and [[mu].sub.0] = 0 for the first series only. Figures 2B and 3B show both estimated latent variables, along with the thresholds associated with each survey. The resemblance to those reported in Figures 2A and 3A is clear. The remaining analysis is conducted using these latent variables.

B. Basic Regressions

To determine the macroeconomic underpinnings of these survey responses, we regress the associated latent variables on a set of economic fundamentals. The set we use includes and exceeds those employed in political business cycle models and analyses of the macroeconomics of happiness: inflation, the unemployment rate, output growth, a medium-term interest rate (the 7-year Treasury bill), an index of the strength of the dollar, and a time trend. (3)

The first three variables in this list are not instantaneously reported with perfect accuracy. Preliminary values for unemployment and output growth are reported by the appropriate federal agency with a lag of one month, and then subsequently revised. Also, the CPI, used in calculating inflation, is reported with a 1-month delay. Which, then, should be used: the ex post, revised values, or the "real-time" data available to respondents at the time the survey is taken?

Our answer is: both. Initially, we use the ex post, revised values, which best measure the "true" value of that variable in that month. Then, subsequently, we replace these values with those constructed using real-time data (distributed by the Federal Reserve Banks of Philadelphia and St. Louis). If the public's economic assessments are based mostly on reported statistics, the real-time data should have superior explanatory power. On the other hand, if these assessments stem from individuals' genuine perceptions of macroeconomic conditions, the ex post, revised data should be superior. Means and detrended standard deviations for the ex post data are listed in Table 4.

The five variables measuring economic fundamentals are often characterized by different integration orders. GDP growth is typically found to be stationary, while inflation and interest rates are 7(1); there is much less agreement on the integration properties of the unemployment rate and dollar strength indicator. But, rather than deepen the integration order of our measures, we prefer to look for robustness. Thus, we estimate both level and difference specifications for the "good economy" question, and also regress the "better/worse" latent variable on differences of the independent variables. The levels specification can be viewed as estimating an a priori known cointegration relationship, with the observed autocorrelation in the residuals representing slow adjustment of survey responses to their long-run fundamental level. The differences specifications, which are more appropriate statistically, estimate short-run relationships.

These specifications require a total of three different differences: (1) in output and price levels, to determine output growth and inflation, (2) of all dependent and independent variables in some "good economy" specifications, and (3) of the independent variables in the "better/worse" regression. Our earlier findings suggest only the length of this last difference: 6 to 8 months (we use eight, which maximizes explanatory power). For the other two, we simply take one year backward differences. These nearly maximize explanatory power and, in the differenced "good economy" regressions, also control for seasonality without sacrificing degrees of freedom.

Estimation is conducted using ordinary least squares (OLS), which yields consistent coefficient estimates in all specifications. But OLS standard errors are biased in cointegrating regressions and when, as here, the error term is serially correlated, so these are adjusted using the Newey-West correction. (4) Three sample periods are analyzed: June 1976 to February 2010, the full time span of the "better/worse" question, December 1985 to February 2010, the full time span of the "good economy" question, and December 1985 to August 2008, which eliminates the September 2008 credit crunch and its aftermath.

Estimates are presented in Table 4. Each independent variable's standard deviation is approximately one, and the thresholds separating different response options are usually little more than one unit apart, so, loosely speaking, each coefficient translates a one standard deviation change in the independent variable into the probability the respondent will choose the next most favorable response option. By a wide margin, economic assessments are most strongly influenced by unemployment. Increasing this by 2 or 3 percentage points will cause at least half of all respondents to choose the next worse response option. Significant but smaller effects, in the expected direction, are also observed with inflation and GDP growth. In contrast, the exchange rate and interest rate generally have smaller, variable, and insignificant coefficients, suggesting that they are secondary. These estimates are all reasonably consistent across specifications and sample periods. (5)

The similarity of key coefficients across specifications occurs partly because assessments of the macroeconomy are adjusted almost instantaneously, as indicated by unreported regressions using lagged dependent variables. This similarity implies that, to the first order, the deterministic factor in our regressions is common to both the "good economy" and "better/worse" series. But its relative importance differs: this factor explains over one-half of the variation in the differenced "good economy" measure, but only one-third of the variation in the "better/worse" measure. Because sampling error is so small, almost all of the remaining variation consists of "intangibles"--unmeasured factors, animal spirits (Akerlof and Shiller 2009), and so on. These intangibles are also, to a considerable extent, common to the two series-thus the differenced "good economy" measure, in Table 2, explains far more variation in the "better/worse" latent variable than any of the regressions in Table 4.

The effects of individual variables are depicted more concretely in Figure 4. This contains a continuous decomposition of the contributions of changes in inflation, unemployment, and GDP growth, and the residual, to the value of the "better/worse" latent variable, using the estimates in the last column of Table 4. (The effects of the exchange rate and the interest rate, which are insignificant in this regression, are suppressed for clarity.) These four components are demeaned and stacked: the value of the unemployment component is [[??].sub.U] ([DELTA][U.sub.t] - [bar.[DELTA]U], and so on. Thus, at each point in time, extending upward from the horizontal axis is the cumulative positive contribution of these four factors toward the deviation of the latent variable from its mean, and similarly for the negative contribution.

In this graph, recessions, recoveries, and expansions are all apparent, as is the relative quiescence of the "Great Moderation." Intangibles (represented by the residual) "explain" the most variance in the latent variable, followed by unemployment and GDP growth. The timing of these three factors is also different: the intangibles have a canary in the coal mine quality (see Chauvet and Guo 2003), plummeting before recessions, while unemployment, naturally, tends to lag. There are no obvious perturbations in the deterministic factor or intangibles associated with national elections.

C. Tradeoffs

The findings in Table 4 can be viewed in terms of tradeoffs, combinations of macroeconomic states that the public considers equally acceptable. The most important tradeoff, between unemployment and inflation, occurs in several contexts: in studies of the "macroeconomics of happiness," in creating simple summary measures of economic conditions, and in conducting macroeconomic policy. It is valuable to see how these compare.

Our estimates indicate that, in assessing the macroeconomy, the public values a 1 percentage point decrease in unemployment as much as a decrease in inflation of 2 to 5 percentage points, depending on the specification. Similar tradeoffs occur in simpler regressions, not reported here, that remove all other independent variables except the trend.

How does this compare to happiness studies that relate measures of happiness or life satisfaction to unemployment and inflation? There, the estimates admit a preference-based interpretation, and the implied marginal rate of substitution of inflation for unemployment ranges from 2:1 to 4:1 (DiTella, MacCulloch, and Oswald 2001, 2003; Wolfers 2003; Blanchflower etal. 2013); the effects of long-term interest rates and economic growth are secondary (Oswald 1997; Welsch 2007, 2011). Virtually all of this work relies on European data, though DiTella, MacCulloch, and Oswald (2001) provide comparable (but less precise) estimates for the United States. The similarity of these findings to ours suggests we cannot rule out the possibility that members of the public evaluate the economy according to their preferences over macroeconomic states.

In contrast, these tradeoffs run counter to the simplest common metric of aggregate economic conditions, Okun's "Misery Index," the unweighted sum of unemployment and inflation. In contrast, equal weights cannot be rejected in explaining the Index of Consumer Sentiment or the Consumer Confidence Index, according to Lovell and Tien (2000) and Golinelli and Parigi (2004). But we have shown that the two sets of surveys measure distinct concepts. The Misery Index seemingly serves better as an indicator of consumer confidence than as a metric of aggregate economic conditions.

Finally, we compare measurement with theory: the tradeoffs embodied in models of monetary policy, often derived from an approximation to household utility in an intertemporal maximization problem. The best known of these, Woodford (2003, chapter 6), is expressed in terms of output growth and inflation, which can be translated into an inflation-unemployment tradeoff using Okun's law (Abel and Bemanke 2005; Mitchell and Pearce 2010). Doing so, we find that this loss function places far more weight on inflation than unemployment, in contrast with our findings and those of the macroeconomics of happiness.

IV. THE INFORMATION CONTENT OF ECONOMIC ASSESSMENTS

The deterministic and intangible components of these survey responses could contain novel information about the current or future values of macroeconomic fundamentals, which, as mentioned, are measured with a lag and (often) subsequently revised. The deterministic component could contain such information if it primarily reflects individuals' perceptions of economic conditions, rather than the reporting of economic statistics (as in Dorns and Morin 2004 or Starr 2012). It does. In the "real-time" row of Table 4 we re-estimated the regressions above using real-time data for unemployment, output growth, and inflation, on which reported statistics would be based. The [R.sup.2] values show that these data have much less explanatory power.

Thus, our latent variables should help predict cunent values of macroeconomic fundamentals that are yet to be reported, and possibly future revisions of past values that have already been reported. One candidate fundamental is real GDP, because the surveys ask about the national economy; another, based on the previous section's results, is unemployment. In that spirit, we predict ex-post (superscript EX) revised values of GDP growth ([DELTA]T) and unemployment (U) in month [t.sup.*] from the real-time (superscript RT) macro and survey data available in the middle of month t, as follows:

(2) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

where [DELTA][pi] is inflation, S contains survey data, (6) {[alpha],[gamma],[phi],[psi],[theta]} are coefficient vectors or matrices, q is time in quarters ([DELTA]Y)/3-month intervals (U), and [xi] is an error term. The term [X.sup.RT.sub.t,0] is the most recent reported value of variable X as of time f, [X.sup.RT.sub.t,-1] is the most recent reported value of that variable 3 months (or one quarter) previous, and so on.

Three types of regressions are run, each with different timing. Nowcasts relate final, revised current-period values to current period realtime data: [t.sup.*] = t. Hindcasts relate final, revised, previous-period values to current period realtime data: [t.sup.*] -1 - 1. These determine whether survey data can predict revisions of macro variables. We also estimate forecasts that relate final, revised future values to current-period real-time data: [t.sup.*] > t. Following Bram and Ludvigson (1998) and others, our metric is the adjusted [R.sup.2] statistic, [[bar.R].sup.2]. Mindful of the multidimensionality revealed in Section II, we employ three different combinations of poll data: the "good economy" and "better/worse" latent variables; a consumer sentiment index, either the Index of Consumer Sentiment or the Consumer Confidence Index; or both the latent variables and a sentiment index. This makes it easy to compare the effects of the two types of surveys, and see whether the information they contribute is distinct or overlapping.

The results for real GDP growth are found in the top panel of Table 5. The latent variables and the sentiment indexes improve hindcasts and nowcasts, raising [[bar.R].sup.2] by about 10 percentage points. The information they contain is overlapping: the [[bar.R].sup.2] values are no higher when the latent variables and sentiment indexes are jointly included in the regression. Both also help hind-cast or nowcast unemployment, in the second panel of the table, but the improvement in fit is quite marginal.

Past studies (Matsusaka and Sbordone 1995 and Golinelli and Parigi 2004, each conducted before the advent of real-time data) have found consumer sentiment indexes to be predictive of future GDP growth, so forecasting equations for GDP and unemployment are also presented. Forecasting power must come from the intangible component of the data, which could predict future economic performance in two ways. A belief that the economy is improving could cause increased economic activity. Or surveys could "anticipate" changes in economic activity that are not predictable from past values of macro variables. The results show that our latent variables do indeed provide substantial new information--unlike the sentiment indexes.

The "good economy" and "better/worse" responses are contemporaneously available (now daily, from Gallup), while the consumer sentiment indexes are only released near the end of the month. (7) Giannone, Reichlin, and Small (2005) show that this timeliness matters, and so it is here. In the bottom row of each panel we re-estimate using the "concurrent" sentiment index, that is, treating that index as if it was available in the middle of the month that the survey was taken. For the Michigan index, at least, the fit immediately improves to match that of the latent variables. The informational advantage of the "good economy" and "better/worse" surveys here extends solely from their earlier reportage, not from the content differences identified in the principal component analysis.

Interestingly, this is not so for growth in personal consumption expenditures (PCE), which has been heavily studied using consumer sentiment indexes. The most visible studies, by Carroll, Fuhrer, and Wilcox (1994) and Bram and Ludvigson (1998), found that these indexes do predict future consumption growth, but these studies did not employ real-time data. When Croushore (2005) replicated these studies using real-time data, he found that the consumer sentiment indexes had little value.

Croushore's preferred specification related final, revised values of real PCE growth to four one-quarter lags of real-time PCE growth, the growth in real stock prices, and the sentiment indexes. We use a similar specification, trimming the lags of all survey data to two, as earlier lags turn out to be superfluous. We use monthly data, as opposed to Croushore's quarterly data, replacing quarter lags with 3-month lags. (8) Importantly, we also use the concurrent values of the sentiment indexes, setting aside differences in reporting dates and focusing on the difference in informational content.

The results are presented in the final panel of Table 5. In the hindcast column, the [[bar.R].sup.2] s are all similar: no survey contributes information. For the nowcasts, both do: [[bar.R].sup.2] increases substantially, and by a similar amount, whether our latent variables, the Michigan consumer sentiment index, or both are included. For the forecasts, the value of the latent variables waxes while that of consumer sentiment wanes. The latent variables increase fit by an astonishing 29 percentage points at a four-quarter horizon, while the contribution of consumer sentiment is far smaller or nil (as in Croushore 2005).

Thus, ironically, the simple "good economy" and "better/worse" surveys predict consumption growth far better than the more sophisticated consumer sentiment indexes that have been designed for this purpose. As these surveys do not ask directly about consumption, their predictive power probably stems from intangibles: the effect of economic optimism on consumption choices. Because Section II showed that the "better/worse" question--which has by far the most predictive power here--is not forward-looking, the most plausible mechanism is that optimism causes, rather than anticipates, future consumption growth (as Matsusaka and Sbordone 1995 found for GDP growth).

In summary, the "good economy" and "better/worse" surveys contain useful information about the past, current, and future evolution of fundamental macroeconomic variables. Some of this information is duplicated by consumer sentiment indexes, with a reporting lag; some is not.

V. CONCLUSION

What makes a good economy? A strong labor market, predominantly, though the public also values lower inflation, more economic growth, and a stronger dollar. Changes in these fundamentals also help explain whether the public views the economy as getting better or getting worse. Still, responses to both types of survey questions are strongly influenced by intangibles that have no obvious economic correlate.

The phrasing of the "good economy" question and response options differs across three surveys, engendering differences in raw response probabilities; nonetheless, these responses are governed by the same underlying latent variable. This latent variable is distinct from various consumer sentiment indices published by the University of Michigan and the Conference Board, though all of these surveys exhibit cyclicality. The same is true for the "better/worse" question, which looks backward with hindsight of 6 to 8 months.

Responses to the "good economy" and "better/worse" questions are based primarily on respondents' perceptions of economic conditions, not media reports of fundamental economic variables. Partly because of this, these responses contain new, timely information about the past, current, and future values of consumption growth, GDP growth, and unemployment. The recently expanded, daily tracking of these two questions by Gallup therefore promises to be a valuable source of information about economic fundamentals.

ABBREVIATIONS

CPI: Consumer Price Index

GDP: Gross Domestic Product

GLS: Generalized Least Squares

ICPSR: Interuniversity Consortium for Political and Social Research

NYT: New York Times

OLS: Ordinary Least Squares

PCE: Personal Consumption Expenditures

doi: 10.1111/ecin.12085

REFERENCES

Abel, A., and B. Bernanke. Macroeconomics. 5th ed. Upper Saddle River, NJ: Pearson, 2005.

Akerlof, G., and R. Shiller. Animal Spirits: How Human Psychology Drives the Economy, and Why It Matters for Global Capitalism. Princeton, NJ: Princeton University Press, 2009.

Bai, J. "Inferential Theory for Factor Models of Large Dimensions." Econometrica. 71, 2003, 135-71.

Bewley, T. "Interviews as a Valid Empirical Tool in Economics." Journal of Socio-Economics, 31, 2002, 343-53.

Blanchflower, D., D. Bell, A. Montagnoli, and M. Miro. "The Effects of Macroeconomic Shocks on Well-Being." Manuscript, University of Stirling, 2013.

de Boor, C. A Practical Guide to Splines. New York: Springer Verlag, 1978.

Bram, J., and S. Ludvigson. "Does Consumer Confidence Forecast Household Expenditure? A Sentiment Index Horse Race." Economic Policy Review, 4(2), 1998, 59-78.

Carroll, C., J. Fuhrer, and D. Wilcox. "Does Consumer Sentiment Forecast Household Spending? If So, Why?" American Economic Review, 84(5), 1994, 1397-408.

Chauvet, M., and J.-T. Guo. "Sunspots, Animal Spirits, and Economic Fluctuations." Macroeconomic Dynamics, 7(1), 2003, 140-69.

Croushore, D. "Do Consumer Confidence Indexes Help Forecast Consumer Spending in Real Time?" North American Journal of Economics and Finance, 16. 2005, 435-50.

DiTella, R.. R. MacCulloch, and A. Oswald. "Preferences Over Inflation and Unemployment: Evidence from Surveys of Happiness." American Economic Review, (91), 2001, 335-41.

---. "The Macroeconomics of Happiness." Review of Economics and Statistics, 85, 2003, 809-27.

Doms, M., and N. Morin. "Consumer Sentiment, the Economy, and the News Media." Federal Reserve Board Working Paper, 2004.

Easaw, J., and A. Ghoshray. "News and Households' Subjective Macroeconomic Expectations." Journal of Macroeconomics, 32(1), 2010, 469-75.

Giannonne, D., L. Reichlin, and D. Small. "Nowcasting GDP and Inflation: The Real-Time Informational Content of Macroeconomic Data Releases." Federal Reserve Board, 2005.

Golinelli, R., and G. Parigi. "Consumer Sentiment and Economic Activity: A Cross Country Comparison." Journal of Business Cycle Measurement and Analysis, 1(2), 2004, 147-70.

Gordon, R. Macroeconomics. 10th ed. Upper Saddle River, NJ: Pearson, 2006.

Litwin, M. How to Measure Survey Reliability and Validity. New York: Sage, 1995.

Lovell, M. C., and P. Tien. "Economic Discomfort and Consumer Sentiment." Eastern Economic Journal, 26(1), 2000, 1-8.

Lux, T. "Rational Forecasts or Social Opinion Dynamics: Identification of Interaction Effects in a Business Climate Survey." Journal of Economic Behavior and Organization, 72(2), 2009, 638-55.

Manski. C. "Measuring Expectations." Econometrica, 72(5), 2004, 1329-76.

Matsusaka, J., and A. Sbordone. "Consumer Confidence and Economic Fluctuations." Economic Inquiry, 36(2), 1995, 296-318.

Merkle, D., G. Langer, and D. Sussman. "Consumer Confidence: Measurement and Meaning." Manuscript, ABC News, 2004. http://www.abcnews.go.com/images/ PollingUnit/ConsumerConfidence061804.pdf.

Mitchell, K., and D. Pearce. "Do Wall Street Economists Believe in Okun's Law and the Taylor Rule?" Journal of Economics and Finance, 34(2), 2010, 196-217.

Oswald, A. "Happiness and Economic Performance." Economic Journal, 107(5), 1997, 1815-31.

Starr, M. "Consumption, Sentiment, and Economic News." Economic Inquiry, 50(4), 2012, 1097-111.

Takezawa, K. Introduction to Nonparametric Regression. Hoboken, NJ: John Wiley, 2006.

Wasserman, L. All of Nonparametric Statistics. New York: Springer, 2006.

Welsch, H. "Macroeconomics and Life Satisfaction: Revisiting the 'Misery Index'." Journal of Applied Economics, 10(12), 2007, 237-51.

--. "The Magic Triangle of Macroeconomics: How Do European Countries Score?" Oxford Economic Papers, 63(1), 2011, 71-93.

Wolfers, J. "Is Business Cycle Volatility Costly? Evidence from Surveys of Subjective Well-Being." International Finance, 6(1), 2003, 1-31.

Woodford, M. Interest and Prices: Foundations of a Theory of Monetary Policy. Princeton, NJ: Princeton University Press, 2003.

(1.) Merkle, Langer, and Sussman (2004) detail the differences in timing, sampling, sample sizes, and interview methods across the ABC News, Michigan, and Conference Board surveys.

(2.) The coefficients on the leads are biased upward if positive assessments of the current change in economic conditions favorably affect the economy's performance in the future. This reinforces the conclusion that the better/worse series compare the present to the past, not expectations of the future to the present. Estimations in which the leads were instrumented with lags of the "better/worse" and/or "good economy" variables, though much less precise, also support this conclusion.

(3.) Inflation is calculated using the all-urban Consumer Price Index (CPI), and the unemployment rate is seasonally adjusted. Each quarterly observation of the real, chain-weighted, seasonally adjusted GDP is assumed to pertain to the middle month of each quarter; the other months are calculated by linear interpolation. The trade-weighted index of exchange rates of the United States's most important trading partners, from the Federal Reserve Bank of St. Louis, has been divided by 10 here so that its variation is comparable to that of the other variables.

Additional regressions not reported here "scaled" unemployment, inflation, or both, for expectations. Unanticipated inflation was calculated using expectations from the well-known Livingston survey; this variable performs slightly worse than simple inflation does. Unemployment was adjusted by calculating its deviation from the Natural Rate of Unemployment, taken from Gordon's (2006) macroeconomics text, or replaced with personal income growth, neither to any effect.

(4.) In consecutive months autocorrelation is at least 0.85 in all specifications. Unfortunately, trying to improve estimation efficiency by using generalized least squares (GLS) is impractical: it generates substantial errors in variables bias in the coefficient estimates, because it implicitly differentiates the data a second time, reducing the signal and amplifying the noise. For example, the standard deviation of the sampling and truncation error in the doubly differenced unemployment rate is 0.16 percentage points (www.bls.gov/cps/eetech.methods.pdf), while the standard deviation of this variable in the data is 0.194 percentage points. Thus, the GLS coefficient estimate should be--and is--attenuated by 0.162/0.1942, to one-third of its true value. Similar reductions were observed for other variables.

(5.) The most noticeable difference is a sizeable reduction in the unemployment coefficient, and explanatory power, when the September 2008 to February 2010 period is added to the sample. The credit crunch sharply and immediately reduced optimism about the state of the economy, while the corresponding increase in unemployment was much more gradual.

(6.) The smoothing involved in the construction of the latent variables in Section III is both forward and backward in time. This is obviously problematic for these forecasting regressions. Thus, both latent variables were reconstructed by replacing the splines in Equation (1) with a set of year x month dummy variables. The resulting latent variables are wholly contemporaneous, but are not available for every month in the sample period, reducing the number of observations available for analysis.

(7.) This slightly overstates and understates the case. A much noisier, preliminary value of the Michigan index is

available mid-month, while the final value of the Conference Board index is not available until the end of the month following the survey. See Merkle, Langer, and Sussman (2004).

(8.) Real-time data on the personal consumption expenditures deflator is not available on a monthly basis, so stock prices were deflated with the CPI instead.

DARREN GRANT, I am grateful to Mishuk Chowdhury, Mitchell Graff, Sohna Jaye, Mohammed Khan, Ryan Murphy, Tino Sonora, Mark Tuttle, Charles Vogel, and Jadrian Wooten, who all provided assistance for this project, and to the Washington Post for providing ABC News /Washington Post survey data through 2005. Comments from Jae Won Lee, Vance Ginn, Andrew Oswald, anonymous referees (who contributed some phrasing used herein), Co-Editors Nezih Guner and Gian Luca Clementi, and participants at the Western Economic Association meetings, the Southern Economic Association meetings, and a seminar at Sam Houston State University are also appreciated. The latent variables created herein can be obtained by contacting the author. Color versions of all figures are available in the online version of this article. Grant: Associate Professor of Economics, Department of

Economics and International Business, Sam Houston State University, Huntsville, TX 77341. Phone 1-936294-4324, Fax 936-294-2417, E-mail dgrant@shsu.edu

TABLE 1

Survey Details

Survey Title       Question(s) Asked of
and Sponsor        Respondents

ABC News (part     "Would you describe the
of the Consumer    state of the nation's
Comfort Index):    economy these days as
Levels (a)         excellent, good, not so
                   good, or poor?"

ABC News/          "Do you think the national
Washington Post:   economy is getting
Changes (b)        better, getting worse, or
                   staying about the same?"

New York Times/    "How would you rate the
CBS News Poll:     condition of the national
Levels (c)         economy these days? Is it
                   very good, fairly good,
                   fairly bad, or very bad?"

New York Times/    "Do you think the economy
CBS News Poll:     is getting better, getting
Changes (c)        worse, or staying about
                   the same?"

USA Today/Gallup   "How would you rate
Poll: Levels (d)   economic conditions in
                   this country today--as
                   excellent, good, only fair,
                   or poor?"

USA Today/Gallup   "Right now, do you think
Poll: Changes      that economic conditions
(d)                in the country as a whole
                   are getting better or
                   getting worse?" (the
                   percent volunteering the
                   response "same" also
                   reported)

University of      "Would you say that you
Michigan Index     are better off or worse off
of Current         financially than you were
Economic           a year ago?" and
Conditions (e)     "Generally speaking, do
                   you think now is a good
                   or bad time for people to
                   buy major household
                   items?"

Conference         "How would you rate
Board Present      present general business
Situation          conditions in your
Index (d)          area--good, normal, or
                   bad?" and "What would
                   you say about available
                   jobs in your area right
                   now--plentiful, not so
                   many, or hard to get?"

                   Temporal Span, Survey
                   Frequency, and Response
Survey Title       Reporting                         Observations,
and Sponsor        (Phone Survey Unless Noted)       Sample Period

ABC News (part     Weekly from December 1985 to      291 Observations
of the Consumer    February 2010, nationwide.        in the 291 months
Comfort Index):    Percentages in each category      from December
Levels (a)         are reported for about 1.000      1985 to February
                   respondents over the previous 4   2010
                   weeks. The values reported in
                   the last survey of the month
                   are used.

ABC News/          Reported in 49 of the 110         Reporting too
Washington Post:   months between September 1981     sporadic to be
Changes (b)        and October 1990, and             individually
                   sporadically afterward,           analyzed
                   nationwide. Percentages in each
                   category are reported for about
                   1,000 respondents over the
                   previous week.

New York Times/    October 1987 to present, at       161 Observations
CBS News Poll:     irregular intervals,              in the 268 months
Levels (c)         nationwide. Percentages in each   from October 1987
                   category are reported for least   to February 2010
                   1,000 respondents over the
                   previous three or five days.
                   All surveys in taken in any
                   given month are averaged.

New York Times/    September 1976 to present,        154 Observations
CBS News Poll:     irregular intervals,              in the 235 months
Changes (c)        nationwide. Percentages in each   from August 1990
                   category are reported for at      to February 2010
                   least 1,000 respondents over      (f)
                   the previous 3 to 5 days. All
                   surveys in each month are
                   averaged.

USA Today/Gallup   January 1997 to October 2008,     111 Observations
Poll: Levels (d)   at irregular intervals,           in the 142 months
                   nationwide. Percentages in each   from January 1997
                   category are reported for at      to October 2008
                   least 1,000 respondents over
                   the previous 3 to 5 days. All
                   surveys in each month are
                   averaged.

USA Today/Gallup   February 1997 to October 2008,    109 Observations
Poll: Changes      at irregular intervals,           in the 141 months
(d)                nationwide. Percentages in each   from February 1997
                   category are reported for at      to October 2008
                   least 1,000 respondents over
                   the previous 3 to 5 days. All
                   surveys in each month are
                   averaged.

University of      Monthly, January 1978 to          291 Observations
Michigan Index     present; three or four times      in the 291 months
of Current         yearly, 1951 to December 1977,    from December
Economic           continental U.S. At least 500     1985 to February
Conditions (e)     respondents over the course of    2010
                   the month. Percentage responses
                   (with one decimal place) to
                   each question are available
                   online. The fraction of
                   positive minus negative
                   responses for each question is
                   calculated, averaged, and
                   indexed to 1966:1.

Conference         Monthly, June 1977 to present;    291 Observations
Board Present      bi-monthly, February 1967 to      in the 291 months
Situation          April 1977, nationwide.           from December
Index (d)          Throughout the month about        1985 to February
                   3,500 respond to a mailing of     2010
                   5,000 surveys, made at the end
                   of the previous month. The
                   fraction of all non-neutral
                   responses that are positive is
                   calculated and indexed to 1985.

(a) Available from 2004 forward from: http://abcnews.go.com/Polling
Unit/CCI/, with earlier values obtained via request to the Washington
Post in 2005, when they were a partner with ABC News for the Consumer
Comfort Survey.

(b) Available from the Interuniversity Consortium for Political
and Social Research (ICPSR).

(c) Available from: http://www.cbsnews.com/stories/2007/10/12/
politics/main3362530.shtml.

(d) Data available, from 1997 forward, from: http://www.
pollingreport.com/consumer.htm.

(e) Breakdowns of response percentages by demographic group
available from: http://www.sca.isr.umich.edu/subset/.

(f) Data prior to August 1990 reported too sporadically
to be individually analyzed.

TABLE 2

Regressions of "Better/Worse" Measures on Leads/Lags of
"Good Economy" Positive Responses (Coefficient Estimates, with
Standard Errors in Parentheses)

                                   Percent Saying the Economy
                                      Is "Getting Better"

Percent Saying the                                CBS/NYT
Economy Is "Excellent"                         (Best 2-Month
or "Good" in ABC Survey             CBS/NYT    Combination)

Six-month lead                      -0.01
                                    (0.12)
Four-month lead                      0.20
                                    (0.17)
Two-month lead                       0.18
                                    (0.16)
One-month lead                                     0.61 *
                                                  (0.05)
Current month                        0.37 *
                                    (0.15)
Two-month lag                        0.08
                                    (0.14)
Four-month lag                      -0.39 *
                                    (0.16)
Five-month lag

Six-month lag                       -0.15
                                    (0.16)
Seven-month lag                                   -0.58 *
                                                  (0.05)
Eight-month lag                     -0.12
                                    (0.14)
Ten-month lag                       -0.06
                                    (0.15)
Twelve-month lag                     0.01
                                    (0.11)
E-statistic on null that             6.00          1.04
  coefficient sum is 0 (p value)    (0.02)        (0.31)
[R.sup.2]/Adjusted [R.sup.2]       0.54/0.51     0.49/0.48

                                   Percent Saying the Economy
                                      Is "Getting Better"

Percent Saying the                                USA Today
Economy Is "Excellent"             USA Today/   (Best 2-Month
or "Good" in ABC Survey              Gallup     Combination)

Six-month lead                       -0.10
                                     (0.14)
Four-month lead                       0.12
                                     (0.19)
Two-month lead                        0.29
                                     (0.18)
One-month lead                                      0.86 *
                                                   (0.09)
Current month                         0.76 *
                                     (0.19)
Two-month lag                        -0.07
                                     (0.19)
Four-month lag                       -0.53 *
                                     (0.19)
Five-month lag                                     -0.76 *
                                                   (0.08)
Six-month lag                        -0.14
                                     (0.19)
Seven-month lag

Eight-month lag                      -0.19
                                     (0.18)
Ten-month lag                        -0.25
                                     (0.18)
Twelve-month lag                      0.20
                                     (0.13)
E-statistic on null that              1.47          3.47
  coefficient sum is 0 (p value)     (0.23)        (0.07)
[R.sup.2]/Adjusted [R.sup.2]       0.75/0.72      0.71/0.70

Note: Time trend also included. N = 144 for
CBS/NYT survey and N= 109 for USA Today/Gallup.

* p < .05.

TABLE 3
Correlations and Principal Components Factor Loadings

                           "Good Economy"
                           Latent Variables   Current Indices

                                      USA/               Conference
Variable                   NYT/CBS   Gallup   Michigan     Board

ABC News Good Economy        0.95     0.91      0.72        0.76
                            (222)    (127)     (283)       (283)
NYT/CBS Good Economy                  0.94      0.72        0.71
                                     (127)     (222)       (222)
USA/Gallup Good Economy                         0.67        0.67
                                               (127)       (127)
Michigan Current                                            0.63
                                                           (283)
Conference Board Current

Michigan Expectations

Conference Board
  Expectations
NYT/CBS Better/Worse

USA/Gallup Better/Worse
Joint variance explained

                                                   "Better/Worse"
                           Expectations Indices    Latent Variables

                                      Conference              USA/
Variable                   Michigan     Board      NYT/CBS   Gallup

ABC News Good Economy        0.63        0.58        0.82      0.69
                            (283)       (283)       (269)     (130)
NYT/CBS Good Economy         0.63        0.62        0.81      0.68
                            (222)       (222)       (222)     (130)
USA/Gallup Good Economy      0.57        0.57        0.70      0.65
                            (127)       (127)       (127)     (127)
Michigan Current             0.69        0.67        0.64      0.58
                            (283)       (283)       (269)     (130)
Conference Board Current     0.39        0.42        0.58      0.44
                            (283)       (283)       (269)     (130)
Michigan Expectations                    0.85        0.56      0.52
                                        (283)       (269)     (130)
Conference Board                                     0.54      0.52
  Expectations                                      (269)     (128)
NYT/CBS Better/Worse                                           0.95
                                                              (130)
USA/Gallup Better/Worse
Joint variance explained

                           Principal Component Factor
                           Loadings (N = 127)

Variable                   1st P.C.   2nd P.C.   3rd P.C.   4th P.C.

ABC News Good Economy        0.37      -0.24       0.15      -0.25
NYT/CBS Good Economy         0.37      -0.19       0.09      -0.42
USA/Gallup Good Economy      0.37      -0.18       0.10      -0.43
Michigan Current             0.34       0.16       0.18       0.40
Conference Board Current     0.29      -0.31       0.50       0.56
Michigan Expectations        0.30       0.60       0.06      -0.13
Conference Board             0.30       0.59       0.05       0.04
  Expectations
NYT/CBS Better/Worse         0.35      -0.15      -0.49       0.18
USA/Gallup Better/Worse      0.31      -0.10      -0.64       0.22
Joint variance explained     72%        11%         8%         4%

Notes: All variables are scaled to have the same variance for the
principal component analysis, as is standard. Consistent with the
findings in Table 2, 8-month differences are taken for all "good
economy" latent variables, Michigan indices, and Conference Board
indices.

TABLE 4
Regression Results (Coefficient Estimates,
with Robust Standard Errors in Parentheses)

                               "Good Economy" Latent Variable

                                                    Levels
                     Mean, Standard
                     December 1985    December 1985    December 1985
Independent           to February     to August 2008    to February
Variable                  2010          (N = 274)      2010 (N = 291)

Unemployment              5.74            -0.39 *          -0.23 *
  (percentage            (1.23)           (0.03)           (0.05)
  points)
One-year output           2.63             0.06 *           0.08 *
  growth (percent)        0.67)           (0.02)           (0.02)
Twelve-month              2.88            -0.08 *          -0.11 *
  inflation              (1.20)           (0.02)           (0.02)
  (percent)
Exchange rate             9.10             0.13 *           0.15 *
  (Fed series,           (0.88)           (0.02)           (0.03)
  scaled by 0.1)
Seven-year T-bill         5.85             0.00             0.12 *
  Rate (percentage       (0.78)           (0.02)           (0.04)
  points)
Trend in years             --             -0.02 *           0.02
                                          (0.01)           (0.01)
[R.sup.2]                  --              0.91             0.87
[R.sup.2] using            --              0.89             0.82
  real-time data

                     "Good Economy" Latent Variable

                                   Differences

                     December 1986    December 1986
Independent          to August 2008    to February
Variable               (N = 262)      2010 (N = 279)

Unemployment             -0.36 *          -0.19 *
  (percentage            (0.05)           (0.06)
  points)
One-year output           0.03             0.05
  growth (percent)       (0.02)           (0.03)
Twelve-month             -0.06 *          -0.07 *
  inflation              (0.02)           (0.01)
  (percent)
Exchange rate             0.08 *           0.06 *
  (Fed series,           (0.03)           (0.03)
  scaled by 0.1)
Seven-year T-bill         0.02             0.08 *
  Rate (percentage       (0.02)           (0.03)
  points)
Trend in years             --               --

[R.sup.2]                 0.69             0.54
[R.sup.2] using           0.68             0.45
  real-time data

                               "Better/Worse" Latent Variables

                                    (On Differences of)

                     December 1985    December 1985    June 1976 to
Independent          to August 2008    to February     February 2010
Variable               (N = 274)      2010 (N = 291)     (N = 405)

Unemployment             -0.32 *          -0.17 *         -0.25 *
  (percentage            (0.08)           (0.08)          (0.09)
  points)
One-year output           0.05             0.08 *          0.06 *
  growth (percent)       (0.03)           (0.03)          (0.02)
Twelve-month             -0.09            -0.10 *         -0.08 *
  inflation              (0.05)           (0.04)          (0.04)
  (percent)
Exchange rate             0.08             0.04            0.05
  (Fed series,           (0.06)           (0.06)          (0.06)
  scaled by 0.1)
Seven-year T-bill        -0.03             0.01           -0.10
  Rate (percentage       (0.05)           (0.05)          (0.05)
  points)
Trend in years             --               --              --

[R.sup.2]                 0.32             0.27            0.36
[R.sup.2] using           0.26             0.14            0.24
  real-time data

Notes: Final revised values of each independent variable are used
in the regressions reported in the table. The [R.sup.2] values
for regressions using real/time data, instead, are reported in
the last row. Each regression also includes a constant. As
discussed in the text, differences are taken over 12 months for
the "good economy" regressions and over 8 months for the "better/
worse" regressions.

* p < .05

TABLE 5

Nowcasting and Forecasting Exercises, Using Real-Time
Data for All Independent Variables (Adjusted [R.sup.2] Values)

                             Index: Michigan Index
                             of Consumer Sentiment

                                             IQ         4Q
Dependent Variable   Hindcast   Nowcast   Forecast   Forecast

Real GDP growth
  No survey data     0.397      0.316     0.280      0.453
  Latent variables   0.483 *    0.411 *   0.390 *    0.533 *
  Sentiment index    0.478 *    0.433 *   0.315 *    0.476
  Both               0.485 *    0.438 *   0.387 *    0.530 *
  Concurrent index      --      0.427 *   0.370 *    0.498 *
Unemployment rate
  No survey data     0.997      0.985     0.939      0.667
  Latent variables   0.997 *    0.987 *   0.955 *    0.763 *
  Sentiment index    0.997 *    0.987 *   0.952 *    0.729 *
  Both               0.997 *    0.987 *   0.957 *    0.766 *
  Concurrent index      --      0.987 *   0.955 *    0.761 *
Real PCE growth
  No survey data     0.692      0.112     0.198      0.149
  Latent variables   0.692      0.275 *   0.318 *    0.437 *
  Concurrent index   0.697      0.289 *   0.254 *    0.264 *
  Both               0.701      0.294 *   0.310 *    0.450 *

                             Index: Conference Board
                            Consumer Confidence Index

                                             IQ         4Q
Dependent Variable   Hindcast   Nowcast   Forecast   Forecast

Real GDP growth
  No survey data     0.397      0.316     0.280      0.453
  Latent variables   0.483 *    0.411 *   0.390 *    0.533 *
  Sentiment index    0.480 *    0.392 *   0.299      0.463
  Both               0.494 *    0.420 *   0.387 *    0.532 *
  Concurrent index      --      0.437 *   0.327 *    0.475
Unemployment rate
  No survey data     0.997      0.985     0.939      0.667
  Latent variables   0.997 *    0.987 *   0.955 *    0.763 *
  Sentiment index    0.997 *    0.987 *   0.952 *    0.718 *
  Both               0.997 *    0.987 *   0.956 *    0.774 *
  Concurrent index      --      0.987 *   0.956 *    0.754 *
Real PCE growth
  No survey data     0.692      0.112     0.198      0.149
  Latent variables   0.692      0.275 *   0.318 *    0.437 *
  Concurrent index   0.690      0.146 *   0.200      0.175 *
  Both               0.691      0.319 *   0.315 *    0.443 *

Note: N = 215.

* Joint significance of the survey data included,
relative to the baseline of no survey data.