Categories
American Economist

Can Consumer Attitudes Forecast Tile Macroeconomy?

Can Consumer Attitudes Forecast Tile Macroeconomy?

Marc C. Chopin

Marc C. Chopin [*]

Ali F. Darrat [**]

I. Introduction

Researchers and policy-makers have often used surveys of consumer attitudes to forecast economic performance. The perceived importance of consumer attitudes is evidenced by the attention paid to announcements made by the Conference Board and the inclusion of the Index of Consumer Confidence (ICC) in the Commerce Department’s Index of Leading Economic Indicators. However, as Katona (1978) notes, the use of consumer attitudes for forecasting is based on the assumption that “attitudes and expectations intervene between stimuli and response and they change before behavior changes.” If changes in the attitudes precede changes in consumer behavior, then knowledge of these attitudes could help explain consumer spending and saving patterns [Liken and Kotler (1983), Kinsey and Collins (1994)]. However, if attitudes change after or concurrently with other movements in the economy, then measures of consumer attitudes will add little to models designed to forecast the economy. Therefore, the temporal ordering between co nsumer attitudes and their behavior should determine the value of attitudes measures in forecasting models.

Studies examining the value of consumer attitudes for forecasting economic performance have thus far failed to produce a consensus. For example, Juster and Watcher (1972), Kelly (1990), Throop (1991), and Carrol, Fubrer and Wilcox (1994) all report evidence suggesting that the ICC contains useful information for predicting consumer spending. In contrast, Hymans (1970), Lovell (1975), and Burch and Gordon (1984) contend that such measures are poor predictors of consumer spending. In spite of the controversy surrounding the predictive value of consumer attitudes, Gaski and Etzel (1986) conclude that “these surveys (measuring consumer attitudes) are still used in business planning.”

We should point out that most of previous studies in this area focus on the contemporaneous correlation between some measures of consumer attitudes and economic conditions. However, in forecasting models, contemporaneous correlations are of limited usefulness. Indeed, it is now widely recognized that strong correlations between any two variables are insufficient to identify the cause and effect relationships between them. More valuable information may be found by examining movements in variables that precede changes in others. For example, if changes in consumer attitudes precede changes in consumer spending, then measures of consumer attitudes will be useful for economic and business forecasting. However, if changes in consumer spending precede changes in consumer attitudes, then measures of consumer attitudes will have no forecasting value.

Empirical work on the above issue of causality remains extremely sparse. Recently, two papers have investigated the causal relationship between indices of consumer attitudes and other economic variables, and they too report conflicting results. After testing for Granger-causality, Garner (1991) concludes that the ICC (and a similar measure published by the University of Michigan) are largely unreliable predictors of consumer spending. On the other hand, Huth, Eppright, and Taube (1994) also examine the Granger-causal relationships between four measures of consumer attitudes and various business and economic variables. In contrast to the conclusions reached by Garner, Huth et al. claim that the four measures of consumer attitudes provide useful information for forecasting changes in several measures of consumer spending, business and economic activity.

While these two studies represent an important step forward in the examination of the information content of measures of consumer attitudes, both studies appear seriously flawed. In particular, they examine the causal linkage in the context of bivariate models whose inferences are known to be potentially biased due to the “omission-of-variables” phenomenon [Lutkepohl (1982, 1993)]. Furthermore, although Granger-causality tests are sensitive to changes in the lag structure [Thornton and Batten (1985)], both studies are tainted by the fact that they impose a common lag for all variables. More disturbing, perhaps, is that neither study incorporates possible cointegratedness among the variables. If cointegration does in fact exist, then both studies become seriously misspecified, as we will discuss below.

In this paper, we use a flexible lag structure and a multivariate model to investigate the Grangercausal relationships among consumer attitudes and several macro variables, namely, retail sales, personal disposable income, inflation, stock prices, money supply, and interest rates. We also test for the stationarity of the series and check for possible cointegration underlying them. We derive our Granger-causality inferences from a multivariate vector-error-correction-model (VECM).

The remainder of the paper is organized as follows. Section II provides a brief description of the data along with the results from stationarity and cointegration tests. Section III reports the causality results. Section IV offers some concluding remarks.

II. Data, Stationarity and Cointegration Results

The Conference Board’s Index of Consumer Confidence (ICC) is widely used as a measure of consumer attitudes and has received considerable attention in both the academic and popular press. Furthermore, Huth, et. al. (1994) find that the Conference Board’s ICC is preferred to the University of Michigan’s ICS when predicting changes in economic activity. These factors have piqued our interest in the ICC as the measure of consumer attitudes. Our objective is to estimate the marginal value of measures of consumer confidence in models designed to predict macroeconomic activity and changes in retail sales.

Garner (1991), Leeper (1992) and Throop (1991) suggest that, particularly during normal economic times, consumer attitudes merely reflect the state of the economy. For example, according to these researchers, increases in consumer’s income and decreases in inflation or interest rates, ceteris paribus, will result in improved consumer wellbeing and attitudes. If consumer attitudes do nothing more than reflect past or prevailing economic conditions, then measures of consumer confidence will add little or nothing to models including measures of economic performance upon which consumer attitudes are based. In contrast, Katona and Mueller (1953) argue that consumer attitudes are not merely a reflection of current economic conditions and that measures of macroeconomic performance are not viable substitutes for measures of consumer attitudes. This question of whether consumer attitudes simply reflect changes in the economy or contain information not captured by measures of economic performance lies at the heart of the debate about the value of consumer attitudes in models used to predict changes in consumer spending.

To identify the Granger-causal ordering and estimate the marginal impact of measures of consumer confidence on estimates of consumer spending, our model includes several measures of economic performance and policy variables. To incorporate possible effects of consumers’ income and wealth on consumer confidence, we include personal disposable income (PI) and the Dow Jones Industrial Average (DOW) in the model. As we note above, we expect to find a positive relationship between income and consumer confidence. Furthermore, increases in the DOW are expected to reflect market participants forecasts of firm profitability, interest rates and the economy. We expect increases in the DOW to reflect improvements in the economic well being of consumers. To capture the impact of changes in the cost of living and the cost of borrowing on consumer attitudes, we add the CPI annualized inflation rate (INF) and the three-month Treasury bill rate (SR). To reflect the impact of monetary policy on the economy and consumer attitu des, we include changes in the monetary base (MB). Changes in retail sales proxy for changes in consumer spending (RT). Since most economic decisions occur at the microeconomic level, it may be advisable to scale the variables by the size of population when analyzing macroeconomic issues. Therefore, we express all aggregate variables in their per capita figures. [1] The tests are implemented at the monthly frequency over June 1977 to November 1996, and the data came from the DRI/Citibase data tape.

As noted above, we examine the causal relationship between consumer attitudes and several macroeconomic variables by testing for Granger-causality using a multivariate VECM. Prior to estimations, we inspect our variables for any possible nonstationarity. [2] Granger and Newbold (1974) demonstrate that nonstationary time series yield spurious results. Further, Phillips (1986) and Stock and Watson (1989, 1993) find that nonstationary data may produce incorrect inferences since t-, F-and [[chi].sup.2] statistics do not converge to their limiting distributions even asymptotically. After converting our variables to stationary series, we then test for cointegration (long-run relationship) among the variables. Models that ignore possible cointegratedness underlying the variables, when it exists, are biased for they filter out important low-frequency information [See Harris (1995)]. Finally, we allow for a flexible lag structure for each variable, with the order of entry and the number of lags of each variable inclu ded in the model determined using the Akaike FPE procedure in conjunction with the well-known specific-gravity criterion.

Table 1 reports the results from the Augmented Dickey-Fuller (ADF) test of nonstationarity. As we can see from the table, all series appear nonstationary in levels, but become stationary in first-differences. These results are robust to the inclusion or exclusion of a time trend. [3]

Having determined that each of the seven series is stationary in first-differences, we test next for the existence of cointegration (long-run equilibrium) relationships among the variables included in the model. In light of recent econometric literature [e.g., Cheung and Lai (1993), and Gonzalo (1994)], we employ the Johansen (1988) efficient procedure to test for cointegration, and also to identify the number of cointegrating vectors, should cointegration be found. When testing for cointegration, we use the Akaike Information Criterion (AIC) to determine the proper lag length in the Johansen model. Results show that six lags minimize the AIC and also whiten the residuals. We note that all seven variables were initially included in the cointegrating space under investigation. However, a likelihood ratio (LR) test suggested that one of the variables (the DOW) should be excluded to yield a parsimonious cointegrating vector [[[chi].sup.2] = 2.39, [[chi].sup.2](0.05) = 7.8l]. [4] The Johansen test results for the parsimonious vector are assembled in Table 2.

The Johansen test statistics are adjusted for the finite-sample bias using the correction factor proposed by Reimers (1992), and Cheung and Lai (1993). Both the trace and the maximal eigenvalue tests of the Johansen approach suggest that there are three non-zero (stationary) cointegrating vectors linking the variables. Thus, estimating a standard VAR without due allowance for the underlying cointegrating relationship will yield biased results. Granger’s (1991) Representation Theorem further implies that we should specify a Vector Error Correction Model, VECM, (instead of a VAR) in which causality is detected among the variables in at least one direction. To ensure statistical efficiency, we estimate error-correction terms from the parsimonious six-variable cointegrating vector using the Johansen maximum-likelihood method. These error correction terms then enter each of the equations prior to estimating the order of entry and the lag structure for the remaining variables in each equation. Our next task is to identify the Granger-causal inter-relationships in the context of the VECM.

III. Causality Linkages Between Consumer Attitudes and the Macroeconomy

As Ahking and Miller (1985) and Thornton and Batten (1985) demonstrate, the use of a common lag for all variables in a given model is overly restrictive and theoretically baseless. In addition, Granger-causality tests and policy inferences from VARs or VECMs are known to be very sensitive to lag specifications [Hafer and Sheehan (1991)]. Therefore, unlike Garner (1991) and Huth et al. (1994), an objective data-based technique should be used to specify the model’s lag structure. According to the Monte Carlo evidence reported by Thornton and Batten (1985), we employ Akaike’s final prediction error (FPE) criterion to select the appropriate lag structure for each variable included in our VECM. [5] Once again, it is important to note that we avoid possible biases resulting from the omission of the underlying long-run relationships among the variables by including the associated error correction terms. [6]

We use the FPE procedure to estimate seven equations, one for each endogenous variable. We then pool the seven resultant equations and estimate them as a system using Zellner’s Seemingly Unrelated Regression (SUR) technique. Zellner’s SUR yields consistent and asymptotically efficient estimates on the assumption that the errors in each equation are themselves uncorrelated. With the final VECM as the maintained hypothesis, we conduct a series of over- and under-fitting tests using SUR system estimations to further refine the model specification. That is, we add and then remove two lags of each variable in each equation, one at a time, and use the likelihood ratio test to determine the significance of the added or deleted lags. Over- and under-fitting tests amount to testing the robustness of the maintained model to alternative lag specifications. We also test for autocorrelation of the residuals and the temporal stability of the coefficient estimates. [7] Finally, we perform likelihood ratio tests within the system estimations to test for the joint significance of the lagged coefficient estimates for each variable in each equation. [8] These system likelihood ratio tests form the basis of our Granger-causality inferences. The final VECM results take the form of model (1). The variables DICC, DRT, DSR, DMB, DINF, DPI, and DDOW represent, respectively, the logarithmic [9] first-differences of the following variables, the index of consumer attitudes, per capita retail sales, the three-month Treasury bill rate, the per capita money base, the annualized CPI inflation rate, per capita disposable personal income, and the Dow Jones Industrial Average. The coefficients, [[[Beta].sub.ij].sup.k](L) denotes the number of lags (k) of variable j included in equation i, [alpha]’s are the constants, [xi]’s are the three error-correction terms, the [gamma]’s are their associated coefficients, and the [e.sub.i]’s are Gaussian disturbance terms.

Table 3 displays the Granger-causality test results on the basis of system likelihood ratio statistics estimated by the SUR procedure. As we mentioned earlier, a large number of over- and under-fitting tests were performed on the maintained model to ascertain the robustness of the specification to various alterations. The results from likelihood ratio tests (relegated to an Appendix available from the authors upon request) generally indicate that the maintained model as given above is an adequate representation of the data during the estimation period.

We now discuss the Granger-causality results. Following Jones and Joulfaian (1991) and Perman (1991), the joint significance of lagged independent variables in any equation reflects short-run Granger-causality, while the significance of the error-correction terms in the equation is indicative of long-run Granger-causality. Since our main concern in this paper is on the predictive content of the variables, we focus below on our findings for short-run linkages that are more relevant to the forecastability issue.

Our results from the VECM reported in Table 3 suggest that only changes in inflation and in the DOW cause significant movements in the measure of consumer attitudes. Of course, this finding is intuitive and can be easily vindicated. Changes in inflation impact the cost of living and the real values of financial assets, both of which are important concerns for consumers. Similarly, changes in the DOW reflect the economy’s overall performance and are linked to consumers’ wealth. Not surprisingly, therefore, these two variables (inflation and the DOW) appear to have significantly shaped consumer attitudes. The results further suggest that the impact of inflation and the DOW on consumer attitudes is relatively swift and completed within two to three months. Perhaps equally important, none of the remaining variables appears to have impacted consumer attitudes. In other words, our empirical results suggest that consumers, once they incorporate information on inflation and the stock market, do not attach much impor tance to changes in other macro variables like interest rates or monetary policy moves when forming their sentiment about the future of the economy.

For forecasting retail sales, we have obtained decidedly strong evidence that only the behavior of the stock market (as proxied by changes in the DOW) causes significant changes in retail sales. In particular, changes in the ICC (and other macro variables) seem to add little to forecast retail sales. We should note that this inference for the ICC is supported by the fact that the FPE criterion suggests the removal of the ICC variables from the retail sales equation. Moreover, to check whether we have been misled by the FPE procedure, we relax this FPE zero restriction by over-fitting the model with two non-zero lags on the ICC variable. The resultant likelihood ratio statistic still fails to indicate significance [[chi].sup.2] = 1.70, [[chi].sup.2](0.05) = 5.99].

In contrast to retail sales, our empirical results indicate that the ICC contains valuable information for predicting personal disposable income [[[chi].sup.2] = 12.86, [[chi].sup.2](0.05) = 7.811. Thus, changes in consumer attitudes appear to anticipate movements in personal income, rather than reflect them. Besides personal disposable income, changes in the ICC also cause significant movements in short-term interest rates [[chi].sup.2] = 9.74, [[chi].sup.2](0.05) = 7.81] and in the DOW [[[chi].sup.2] = 7.06, [[chi].sup.2](0.05) = 5.99]. However, the ICC does not exert any causal impact upon inflation or base money. On balance, then, the ICC appears to be a reliable predictor of certain macro variables, but not of others. To repeat, changes in the measure of consumer attitudes contain useful information for predicting personal disposable income, interest rates and movements in stock prices. On the other hand, the ICC is of little value in models used to forecast retail sales, inflation, or monetary policy. Therefore, unlike Garner (1991) and Huth et. al (1994), the evidence we obtain suggests that measures of consumer attitudes do have predictive contents, though not uniformly across the economic spectrum.

Before concluding, observe that our VECM results can also be useful to distill indirect causality linkages. For example, the ICC does not directly cause changes in retail sales. Yet, it can still influence retail sales indirectly through interactions with the DOW. As the results from the DOW equation indicate, the ICC causes significant changes in the DOW, which in turn impacts retail sales. Thus far, we have focused on short-run causality. However, we should note that each of the seven estimated equations contains three error-correction terms to approximate long-run Granger-causality. With the exception of the DOW equation, at least one of the error-correction terms proves statistically significant in each of the remaining six equations. These results accord well with the presence of a long-run equilibrium relationship among the variables in the parsimonious cointegrating vector. This finding also implies that overlooking cointegratedness would have resulted in biased inferences.

IV. Concluding Remarks

This paper examines empirically the cause-and-effect relationships between consumer attitudes and several key macroeconomic variables. To that end, we use the recent cointegration and error-correction modeling technique in a multivariate setting.

The empirical results we obtain from monthly data over 1977-1996 suggest the presence of significant long-run (equilibrium) relationships among the variables. By virtue of the Granger Representation Theorem, there must be Granger-causality among the variables in the model in at least one direction. Results from our VECM confirm the presence of these causal linkages among several variables in the model.

The evidence which consistently emerges is that the ICC does contain valuable information for predicting some macroeconomic variables but not others. In particular, changes in the ICC can predict movements in personal disposable income, interest rates, and to some extent, also the DOW.

However, the ICC proves an unreliable predictor for retail sales, inflation or monetary policy moves. As for the factors responsible for changes in consumer attitudes, the evidence suggests two possible culprits; namely, changes in inflation and in the DOW. These results lend support to the view that consumers closely watch the inflation outlook and the behavior of the stock market when formulating their own sentiment. Our results partly, although not completely, support the conclusions of Huth et al. (1994). Like them, we find that measures of consumer attitudes are useful for economic forecasting. Unlike them, we do not find evidence that such measures are useful uniformly or even generally across the economic and business spectrum.

(*.) Associate Professor of Economics, Louisiana Tech University

(**.) Professor of Economics, Louisiana Tech University

The authors wish to acknowledge capable research assistance of Thanomsak Suwanoi and Maosen Zhong. Suggestions from an anonymous referee also helped to improve the paper. Any remaining errors are the sole responsibility of the authors.

Notes

(1.) Recent research has also recommended the use of per capita variables in order to avoid measurement errors. See Heston (1994).

(2.) Any variable is said to be stationary if its stochastic properties (mean, variance, and covariance) are time invariant.

(3.) Note that PI proves stationary in first-differences only when a time trend is included in the testing equation.

(4.) The exclusion hypothesis is easily rejected for each of the remaining six variables. The LR statistics (with d.f. = 3) are 24.79 for ICC; 21.05 for retail sales; 219.47 for interest rates; 17.74 for base money; 15.72 for inflation; and 19.61 for personal disposable income.

(5.) The FPE procedure amounts to minimizing a function of the one-step-ahead prediction error. It is a compromise between the predictive power of a model and its complexity, the latter measured by the model order. See Darrat (1988) for a detailed account of the FPE procedure.

(6.) Similar to the bias from omitting one relevant long-run cointegrating relationship from the model, when it exits, the omission of additional relevant cointegrating relationships will also result in an omitted variable bias.

(7.) The results, available from the authors upon request, evince no severe problems of autocorrelations or unstable parameters. Testing for possible pitfalls in our model is important in order to achieve, in the words of Hendry and Ericsson (1991, p. 19), “a congruent model, that is, one which captures the salient features of existing data and is interpretable in the light of available economic theory.”

(8.) Note that the VECM model is specified so that only lagged values of the endogenous variables appear as explanatory variables. Contemporaneous relationships among the variables are reflected in the innovations. [See Hsiao (1981), and Sephton (1988)].

(9.) We apply the natural logarithmic transformation in order to stabilize the variance of the series over time (i.e., induce homoscedasticity). This transformation also appears reasonable since the logarithmic first-differences of a given variable approximate its percentage changes.

References

Ahking, F. W., and S. M. Miller, “The Relationship Between Government Deficits, Money Growth, and Inflation.” Journal of Macroeconomics 7 (Fall 1985): 447-467.

Burch, S. W., and S. Gordon, “The Michigan Surveys and the Demand for Consumers Durables.” Business Economics 19 (October 1984): 40-44.

Carrol, D. C., J. C. Fuhrer, and D. W. Wilcox, “Does Consumer Sentiment Forecast Household Spending? If So, Why?” American Economic Review 84 (December 1994): 1397-1408.

Cheung, Y. W., and K. S. Lai, “Finite-Sample Sizes of Johansen’s Likelihood Ratio Tests for Cointegration,” Oxford Bulletin of Economics and Finance 55 (August 1993): 313-328.

Darrat, A. F., “Have Large Budget Deficits Caused Rising Trade Deficits?” Southern Economic Journal 54 (April 1988): 879-887.

Garner, A., “Forecasting Consumer Spending: Should Economists Pay Attention to Consumer Confidence Surveys?” Federal Reserve Bank of Kansas City, Economic Review 76 (May/June 1991): 57-71.

Gaski, J. F., and M. J. Etzel, “The Index of Consumer Sentiment Toward Marketing.” Journal of Marketing 50 (July 1986): 71-81.

Gonzalo, J., “Five Alternative Methods of Estimating Long-Run Equilibrium Relationships.” Journal of Econometrics 60 (January 1994): 203-233.

Granger, C. W. J., “Developments in the Study of Cointegrated Economic Variables,” Long-Run Economic Relationships: Readings in Cointegration, edited by R. F. Engle and C. W. J. Granger (Oxford: Oxford University Press, 1991): 65-80.

_____, and P. Newbold, “Spurious Regressions in Econometrics.” Journal of Econometrics 2 (July 1974): 111-120.

Hafer, R. W., and R. G. Sheehan, “Policy Influence Using VAR Models.” Economic Inquiry 29 (January 1991): 44-52.

Harris, Richard, I. D., Using Cointegration Analysis in Econometric Modeling (Hertfordshire: Prentice Hall/Harvester Wheatsheaf, 1995)

Hendry, D. F. and N. R. Ericsson, “An Econometric Analysis of U.K. Money Demand in Monetary Trends in the United States and the United Kingdom by Milton Friedman and Anna J. Schwrtz.” American Economic Review 81 (March 1991): 8-38.

Heston, A., “A Brief Review of Some Problems in Using National Accounts Data in Level of Output Comparison and Growth Studies.” Journal of Development Economics 44 (1994): 29-52.

Hsiao, C., “Autoregressive Modeling and Money-Income Causality Detection.” Journal of Monetary Economics 7 (January 1981): 85-106.

Huth, W. L., D. R. Eppright, and P. M. Taube, “The Indices of Consumer Sentiment and Confidence: Leading or Misleading Guides to Future Buyer Behavior.” Journal of Business Research 29 (1994): 199-206.

Hymans, S., “Consumer Durable Spending: Explanation and Prediction.” Brookings Papers on Economic Activity 2 (1970): 173-199.

Johansen, S. “Statistical Analysis of Cointegrating Vectors.” Journal of Economic Dynamics and Control 12 (June-September 1988): 241-254.

Jones, J. D. and D. Joulfain, “Federal Government Expenditures and Revenues in the Early Years of the American Republic: Evidence from 1792 to 1860.” Journal of Macroeconomics 14 (Winter 1991): 133-155.

Juster, F. T., and P. Watchel, “Anticipatory and Objective Models of Durable Goods Demand.” American Economic Review 62 (September 1972): 564-579.

Katona, G., “Behavioral Economics.” Challenge (September/October 1978): 14-18.

_____, and E. Mueller, Consumer Attitudes and Demand, 1950-1952, Ann Arbor, MI: Survey Research Center, University of Michigan.

Kelly, D., “The Plunge in Confidence Will Hit Spending–But How Hard?” DRI/McGraw Hill U.S. Review (September 1990).

Kinsey, J., and M. I. Collins, “Index of Consumer Expectations: Food Price Effects and Durable Goods Expenditures.” The Journal of Consumer Affairs 28 (Winter 1994): 255-277.

Leeper, E. M., “Consumer Attitudes: King for a Day,” Economic Review, Federal Reserve Bank of Atlanta, July/August 1992: 1-15.

Lilien, G., and P. Kotler, Marketing Decision Making. A Model Building Approach (New York: Harper and Row, 1983).

Lovell, M., “Why Was the Consumer Feeling So Sad?” Brookings Papers on Economic Activity 2 (1975): 473-479.

Lutkepohl, H., “Non-Causality Due to Omitted Variables.” Journal of Econometrics 19 (1982): 367-378.

_____. Introduction to Multiple Time Series Analysis, Second Edition (Berlin, Germany: Springer-Verlag, 1993).

Perman, R., “Cointegration: An Introduction to the Literature.” Journal of Economic Studies 18 (1991): 3-30.

Phillips, P. C. B., “Understanding Spurious Regressions.” in Journal of Econometrics 33 (December 1986): 311-340.

Reimers, H. E., “Comparisons of Tests for Multivariate Cointegration.” Statistical Papers 33 (1992): 335-359.

Sephton, P. S., “On Interest Rate Innovations and Anticipated Monetary Policy,” Economics Letters 28 (1988): 177-180.

Stock, J., and M. Watson, “Interpreting the Evidence on Money-Income Causality.” Journal of Econometrics 40 (January 1989): 161-182.

______, “A Simple MLE of Cointegrating Vectors in Higher Order Integrated System.” Econometrica 61 (July 1993): 783-820.

Thornton, D. L., and D. S. Batten, “Lag-Length Selection and Tests of Granger-Causality Between Money and Income.” Journal of Money, Credit and Banking 17 (May 1985): 164-178.

Throop, A., “Consumer Sentiment and the Economic Downturn.” Federal Reserve Bank of San Francisco, Weekly Letter (March 1, 1991).

COPYRIGHT 2000 Omicron Delta Epsilon

COPYRIGHT 2008 Gale, Cengage Learning