The macroeconomic relationship between advertising and consumption.
Profit-maximizing firms advertise in order to increase the demand for the goods they produce. Hence, we would expect increases in a firm's advertising to be associated with increases in consumption of the firm's goods, ceteris paribus. The increased demand for the firm's goods may be associated with decreased demand of their rivals' goods. At the market level, increases in advertising by the industry may not be accompanied by increased consumption of the industry's output. This is because changes in the levels of advertising among firms in a particular industry could merely rearrange market shares. If this is true then, in that industry, a firm which increases its advertising relative to its rivals will gain sales at the rivals' expense; advertising in this case is a zero-sum game.
When we aggregate across markets to consider advertising at the national level, the question of the effect of advertising upon consumption has still different implications. If increased advertising levels for the economy are associated with increased consumption, this suggests that consumers are increasing current consumption at the expense of future consumption (i.e., savings). If this is true, then advertising affects aggregate consumption and the business cycle.
In this paper, we consider the relationship between aggregate advertising and aggregate consumption. While previous research (e.g., Ashley, Granger, and Schmalensee  and Schmalensee ) has suggested that consumption affects advertising, most previous empirical studies of the causal relationship between aggregate advertising and aggregate consumption reject the hypothesis that advertising affects consumption [6, section 8.3; 12]. An early study was Verdon, McConnell, and Roesler , who studied the relationship between advertising and aggregate demand (GNP). While they found advertising to be pro-cyclical and to have a positive effect on aggregate demand, correlations between various leads and lags led them to conclude that no clear pattern existed. This study was followed by Ekelund and Gramm , who faulted the Verdon, McConnell, and Roesler study for considering the relationship between advertising and aggregate demand. Instead, Ekelund and Gramm concentrated upon the relationship between advertising and aggregate consumption, arguing that advertising would not noticeably affect investment and government spending. Nevertheless, Ekelund and Gramm found no relationship between advertising and aggregate consumption.
Taylor and Weiserbs  also examined the relationship between aggregate advertising and aggregate consumption.(1) This study is clearly in the minority; their results suggest that advertising affects aggregate consumption. In a book published the same year, Schmalensee , who was apparently unaware of the then-forthcoming Taylor and Weiserbs study, grappled with the problem of the aggregate advertising-aggregate consumption problem. His results conflict with those of Taylor and Weiserbs by failing to support the hypothesis that advertising affects consumption.
In a still later article, Ashley, Granger, and Schmalensee  faulted the Taylor and Weiserbs study on four grounds: use of the Houthakker-Taylor framework,(2) use of the GNP deflator to deflate advertising expenditures, use of a two-stage least square model with an ad-hoc specification of the advertising equation, and use of annual data. With regard to the latter criticism, they cited Clarke  and Schmalensee [24; chap. 3], both of whom had found advertising effects to depreciate rapidly with most of the advertising effect depreciating within one year. This suggests that annual data may not contain much information about the direction of causation. The results of Ashley, Granger, and Schmalensee  suggest that consumption might affect advertising but that aggregate advertising does not seem to affect aggregate consumption. However, these results were called into question by Berndt  in his review.(3) He pointed out that, while Ashley, Granger, and Schmalensee obtained raw quarterly consumption data, the quarterly 'advertising data they used were seasonally adjusted. Berndt states "it is known that use of seasonally adjusted data can introduce bias into investigations of causality."
Since the Ashley, Granger, and Schmalensee study, which considered the period 1957-75, little research has appeared in the economics literature concerning the aggregate advertising-aggregate consumption question. One study that comes close is Duffy , but he considers the effect of advertising on the distribution of demand across products.(4)
A more closely-related study is Chowdhury , who considers the relationship among advertising and several macroeconomic variables using annual U.K. data over the period 1960-91. While his results support a relationship between advertising and unemployment, they do not support any other relationship including the one between advertising and consumption.
In this paper, we explore the aggregate question using U.S. data and newer techniques than those used in most previous studies. While these techniques are similar to those used by Chowdhury  for his study of the U.K., there are some critical differences from our study, which we discuss in the conclusion.
II. Causal Directions between Advertising and Consumption
In this study, we will perform causality tests between aggregate advertising and aggregate consumption. Before proceeding further with our analysis, we should clarify why one might believe that either one of these aggregates might affect the other.
The hypothesized causal relationship from consumption to advertising is based on the observation that businesses tend to allocate a certain percentage of their (past) revenues to (future) advertising so that increased revenues, which accompany increased consumption, are associated with increased advertising. Some economists (e.g., Hay and Morris ) attribute this to satisfying behavior while others (e.g., Grabowski ) point out how it could reflect profit-maximizing behavior.
The causal relationship from advertising to consumption is credited by Ekelund and Gramm  to Ackley  and Duesenberry . Ekelund and Gramm credit both economists with observing that, if advertising does affect income as claimed by Hansen  (see footnote 1.), it would do so through consumption or, more specifically, through the marginal propensity to consume (MPC). Of course, an increase in the MPC would cause a decrease in the marginal propensity to save (MPS), so the effect of advertising is to increase current consumption at the expense of future consumption. Apart from the effect of advertising on income through the MPC (hence consumption), any effect through investment or government spending would be secondary in nature.
III. The Data
Data for our empirical work were obtained from McCann-Erickson, Inc., of New York City and the Citibank Database. Because unseasonalized quarterly advertising data are unavailable, we use annual data. While it may be true that a portion of the year's advertising will be depreciated and not affect consumption in the subsequent year, this is perhaps not so big a problem as one might think. First, a large portion of any year's advertising expenditures occurs in the fourth quarter of the year, so there is likely to be some remaining effect of last year's advertising on this year's consumption. Second, even if all advertising did occur on January 1 of a year, some of the largest of the depreciation estimates still allow a small proportion of advertising to linger past January 1 of the subsequent year: Schmalensee  suggests that between 15 to 25 percent of the advertising effect will remain after one year and even Clarke  suggests that up to 5 percent will remain. Furthermore, if the remaining advertising effects of the previous year are, in fact, small then any test for a significant effect of last year's advertising on this year's consumption is more stringent than if the remaining effects were large.
McCann-Erickson, Inc. provided total advertising expenditures for 1947-88 in ten millions of dollars. This is deflated using the implicit GNP deflator (1982 = 100) obtained from the Citibank Database. While Ashley et al. constructed an advertising price index, there is evidence that such a construction, which would be subject to its own internal errors, might not contribute much to our analysis. This is because more general deflators often track advertising deflators quite well and can serve as proxies for advertising deflators. For instance, the correlation coefficient between the GNP deflator we use in this analysis and an often-used advertising deflator available from McCann-Erickson of New York for the years 1960-88 is 0.992.(5) Second, there is little difference between results of studies which use an advertising deflator and those which use a more general deflator.(6)
Personal consumption expenditures in billions of 1982 constant dollars (using the consumption component of the GNP deflator) are obtained from the Citibank Database and used as the aggregate consumption variable. The deflator used in constructing this variable is the consumption component of the GNP deflator.
IV. Specification and Cointegration Tests
In order to determine the most appropriate basic functional form for our analysis of aggregate advertising and aggregate consumption, we conduct a specification test to decide whether to specify linear or log-linear models. While a [P.sub.E] test  would be appropriate, it offers no guidance between the linear and the log-linear form. Therefore, we employ a Box-Cox test  for linearity. We calculate a Box-Cox likelihood ratio test statistic ([[Chi].sup.2] with one degree of freedom) equal to 4.68. This is greater than the critical value, 3.84, at the 5 percent level of significance. Therefore, we reject the null hypothesis of linearity at the 5 percent level of significance. Hence, in what follows, we use the log-linear model.
Before we perform a causality test between aggregate advertising and aggregate consumption, we need to test for the stationarity of the two variables or, equivalently, for the presence of a unit root. If variables follow a unit root process, it can lead to spurious results when the levels of the variables are used for estimation purposes because the variance of the process becomes infinite. In that case, least squares estimation with level variables is clearly inappropriate. This problem has traditionally been handled by differencing the data, but this can lead to a significant loss of information. As Granger  points out, however, a set of variables, all of which are stationary only after differencing, may have linear combinations which are stationary without differencing. In such a case the variables are said to be cointegrated.
Engle and Granger  propose a cointegration testing procedure using the Dickey-Fuller  unit root test. However, their cointegration test requires prior knowledge about the cointegrating vectors, which are usually unknown. In contrast, Johansen  and Johansen and Juselius  consider multivariate cointegration tests which can resolve this problem. We employ Johansen's multivariate testing procedure in our analysis.
Before performing a cointegration test, we test for stationarity of the variables by applying the Dickey-Fuller and Augmented Dickey-Fuller tests to determine the order of integration of each variable. The unit root tests are generated by the equation
[Delta][X.sub.t] = [Delta] + ([Rho] - 1)[X.sub.t - 1] + [summation of] [[Theta].sub.i][Delta][X.sub.t - i] where i = 1 to m + [[Epsilon].sub.t] (1)
where [X.sub.t] is a variable, [[Epsilon].sub.t] is the error term, [Delta] is a difference operator, and m indicates the number of differenced terms used. It is the presence of these differenced terms that distinguishes the augmented Dickey-Fuller test from the ordinary Dickey-Fuller test. The Dickey-Fuller test statistic is the ratio of the estimate of [Rho] - 1 to its standard error. The test is conducted under the null [TABULAR DATA OMITTED] hypothesis of a unit root. If the calculated ratio is negative and significantly different from zero using the critical values calculated by Dickey and Fuller , then the null hypothesis is rejected.
Table I shows the results of the Dickey-Fuller and Augmented Dickey-Fuller statistics for each variable. Akaike's  information criterion (hereafter AIC) and the final prediction error (hereafter FPE ) are used to determine the lag order for the augmented terms in equation (1). The unit root test is conducted for both levels and for first differences of each series. According to the results in Table I, the null hypothesis of a unit root is accepted for the level series, but rejected for the first differenced series at the 5 percent level of significance. These results suggest that each variable is nonstationary and integrated of order (1). We can therefore conduct the multivariate cointegration test.
We apply Johansen's likelihood ratio tests to the two variables. As mentioned earlier, Johansen's multivariate tests for cointegration are multivariate extensions of the usual unit root tests for autoregressive processes. Johansen's cointegration tests use canonical correlation of residuals from a reparameterized model to estimate the space of cointegration vectors and to test the dimensions of the space. Consider the p-dimensional process
[X.sub.t] = [[Pi].sub.1][X.sub.t - 1] + [[Pi].sub.2][X.sub.t - 2] + ... + [[Pi].sub.k][X.sub.t - k] + [[Epsilon].sub.t]; t = 1, 2, ... , T, (2)
where [X.sub.t] is a vector of variables with p elements, [[Pi].sub.i] is a p x p matrix (i = 1, 2, ... , k) and [[Epsilon].sub.t] is a vector with p elements composed of independently and normally distributed random disturbances with means 0. An impact matrix can be defined as
[Pi] = I - [[Pi].sub.1] - [[Pi].sub.2] - ... - [[Pi].sub.k], (3)
where all the matrices are p x p and I is the identity matrix. If [Pi] has rank r then there exist r cointegrating vectors in [X.sub.t] and p - r common stochastic trends. We can express the impact matrix as
[Pi] = [Alpha][Beta][prime] (4)
where [Alpha] and [Beta] are matrices of dimension p x r. The space spanned by [Beta] is called the cointegration [TABULAR DATA OMITTED] space and represents the space spanned by the rows of the matrix [Pi]. Even if [X.sub.t] is nonstationary, the linear combination [Beta][prime][X.sub.t] is stationary.
In order to test the hypotheses concerning the number of cointegrating vectors in [X.sub.t], Johansen suggests an estimation of the two equations
[Delta][X.sub.t] = [summation of] [[Theta].sub.1i][Delta][X.sub.t - i] where i = 1 to r - 1 + [[Xi].sub.1t],
[[Xi].sub.t - r] = [summation of] [[Theta].sub.2i][Delta][[Xi].sub.t - i] where i = 1 to r - 1 + [[Xi].sub.2t]. (5)
From these two equations, the vectors of residuals ([[Xi].sub.1t] and [[Xi].sub.2t]) are used to form two likelihood ratio test statistics. The first test statistic, called the trace test [21, 22], is given by the expression
-T [summation of] ln(1 - [[Lambda].sub.i]) where i = r + 1 to p, (6)
where [[Lambda].sub.r + 1], ... , [[Lambda].sub.p] denotes the p - r smallest squared canonical correlations between [[Xi].sub.1t] and [[Xi].sub.2t]. The null and alternative hypotheses for the trace test are [H.sub.0]: r [is less than or equal to] k and [H.sub.A]: r [is greater than] k where, in our case, k = 0 or 1 and r is the number of cointegrating vectors. The second likelihood ratio test, called the maximal eigenvalue test [21, 22], is given by
- T In(1 - [[Lambda].sub.r + 1]). (7)
The null and alternative hypotheses for the maximal eigenvalue test are [H.sub.0]: r = k and [H.sub.A]: r = k + 1, where (again in our case) k = 0 or 1 and r is the number of cointegrating vectors. Johansen and Juselius  suggest that the trace test may lack power relative to the maximal eigenvalue test.
Table II shows the results of the cointegration tests, which we perform in accordance with Johansen and Juselius . Both maximal eigenvalue and trace tests show that there is a single cointegrating relationship between aggregate advertising and aggregate consumption.
V. Error-Correction Model and Granger Causality Test
Engle and Granger  have shown that if two variables are cointegrated, then there exists an error correction model (ECM) for the variables. The cointegrating linear combination of variables is interpreted as an equilibrium relationship, since it can be shown that variables in the error-correction term in an ECM must be cointegrated, and that cointegrated variables must have an ECM representation. Cointegration provides a formal framework for testing for and estimating long-run equilibrium relationships among economic variables. The ECM is a dynamic system in which an error-correction term represents deviations from a long-ran equilibrium relationship while short-run dynamics are represented by lagged difference terms. In general, both the ECM and the vector autoregressive (VAR) model in levels can be used to test for Granger causality. The ECM is, however, more appropriate, since it already embodies the restrictions implied by cointegration and also separates the short-run dynamics and the long-ran equilibrium condition of the variables. Toda and Phillips  compare the Granger causality test in levels VAR and error correction models. They conclude that the causality test based on ECMs performs better than the one based on levels VARs.
In this section, we estimate the following two error-correction models:
[Delta](ln [AD.sub.t]) = [[Alpha].sub.0] + [[Alpha].sub.1][e.sub.t - 1] + [summation of] [[Alpha].sub.2i][Delta](ln [AD.sub.t - i]) where i = 1 to p + [summation of] [[Alpha].sub.3i][Delta](ln [AC.sub.t - i]) where i = 1 to p + [[Eta].sub.1t] (8)
[Delta](ln [AC.sub.t]) = [[Beta].sub.0] + [[Beta].sub.1][e.sub.t - 1] + [summation of] [[Beta].sub.2i][Delta](ln [AC.sub.t - i]) where i = 1 to q + [summation of] [[Beta].sub.3i][Delta](ln [AD.sub.t - i]) where i = 1 to q + [[Eta].sub.2t] (9)
where AC = aggregate consumption, AD = aggregate advertising, [e.sub.t - 1] = ln [AC.sub.t - 1] -[[Delta].sup.e](ln [AD.sub.t - 1]) where [[Delta].sup.e] is the least squares estimate of the parameter in the equation ln [AC.sub.t] = [Delta](ln [AD.sub.t]) + [[Zeta].sub.1], [[Eta].sub.1t], [[Eta].sub.2t], and [[Zeta].sub.t] are the error terms, and, as before, [Delta] is a difference operator.
We estimate equations (8) and (9) for different lag lengths. We determine the optimal lag length by Akalke's AIC and FPE in order to ensure that tests for causality and feedback are not biased by an arbitrary representation of the short-run dynamics. We tried various lag lengths for p and q [is less than or equal to] 12, and found that p = 11 and q = 10 give the minimum of both Akaike's AIC and FPE. Therefore, we estimate the model with p = 11 and q = 10. The estimation results are shown in Table III. In Table III, we note that the coefficients of the error-correction term, [e.sub.t - 1], are negative in the two equations. The error-correction term is interpreted as reflecting disequilibrium responses. The negative signs of the error correction terms assures us that the two equations converge to the long-run equilibrium. If [e.sub.t - 1] is positive (negative) at period t - 1, then [Delta](ln [AC.sub.t]) and [Delta](ln [AD.sub.t]) will be reduced (increased), correcting this error at period t.
Table IV presents the causality testing results. The results shown in Table IV suggest causality with feedback (i.e., two-way causality) consistently, regardless of the lag length used. Even though Akaike's AIC and FPE are minimized at p = 11 and q = 10, we also calculate F-statistics for other lags for which Akaike's AIC and FPE are relatively low because the F-statistic could be sensitive to the choice of lag length. In equation (8), an F-test of the coefficients on the lagged aggregate consumption terms supports the null hypothesis of causality from aggregate consumption to aggregate advertising at the 1 percent level of significance for all the cases considered. In equation (9), an F-test of the coefficients on the lagged aggregate advertising terms also supports the hypothesis of causality from aggregate advertising to aggregate consumption at the 1 percent, 5 percent, or 10 percent levels of significance for all the cases considered.
Table III. Estimation Results (t-statistics in parentheses)
[Delta][AD.sub.t] [Delta][AC.sub.t] Variables Equation (8) Equation (9)
Constant 0.6636 0.1255 (4.7780) (3.0499)
[e.sub.t - 1] -0.1798 -0.0028 (-3.1526) (-0.1378)
[Delta][AC.sub.t - 1] -0.2048 0.4601 (-0.3449) (2.0039)
[Delta][AC.sub.t - 2] -1.4506 -0.2126 (-2.2809) (-2.5098)
[Delta][AC.sub.t - 3] -0.7653 -0.2538 (-1.2172) (-1.0715)
[Delta][AC.sub.t - 4] -0.9547 -0.3417 (-1.4962) (-1.4529)
[Delta][AC.sub.t - 5] -2.4385 -0.4693 (-3.9049) (-2.0244)
[Delta][AC.sub.t - 6] -1.8638 -0.3418 (-2.5938) (-1.3425)
[Delta][AC.sub.t - 7] -2.4105 -0.0635 (-3.5219) (-0.2707)
[Delta][AC.sub.t - 8] -1.9685 -0.4613 (-3.0727) (-2.0209)
[Delta][AC.sub.t - 9] -1.3385 -0.4263 (-1.8739) (-1.6636)
[Delta][AC.sub.t - 10] -1.7957 -0.2484 (-2.6137) (-0.9908)
[Delta][AC.sub.t - 11] -1.3533 - (-2.0430)
[Delta][AD.sub.t - 1] -0.4081 -0.2156 (-2.2504) (-3.0612)
[Delta][AD.sub.t - 2] -0.5572 -0.2126 (-2.5676) (-2.5098)
[Delta][AD.sub.t - 3] -0.7983 -0.1579 (-3.2192) (-1.9120)
[Delta][AD.sub.t - 4] -0.5656 -0.0979 (-2.3442) (-1.2600)
[Delta][AD.sub.t - 5] -0.2524 0.0090 (-1.2101) (0.1168)
[Delta][AD.sub.t - 6] -0.1950 -0.0134 (-0.9985) (-1.3425)
[Delta][AD.sub.t - 7] 0.2535 0.0524 (1.5114) (0.8068)
[Delta][AD.sub.t - 8] 0.0589 0.0447 (0.3479) (0.7063)
[Delta][AD.sub.t -9] 0.0972 0.1842 (0.5858) (3.0420)
[Delta][AD.sub.t - 10] 0.3239 0.0335 (1.7564) (0.4725)
[Delta][AD.sub.t - 11] 0.0584 - (0.3200)
[R.sup.2] 0.8203 0.6780
Durbin-Watson 1.3030 2.1029
[TABULAR DATA OMITTED]
VI. Discussion of Results and Conclusion
In this paper, we have applied error correction models to annual data covering the years 1947-88 to test causality between aggregate advertising and aggregate consumption. As explained above, the use of annual data allows more stringent tests of causality than would quarterly data.
Our results support the hypothesis of two-way causality. More specifically, our results suggest that consumption not only affects advertising, as previous research has shown, but that the converse is also true: aggregate advertising affects aggregate consumption. These results provide some evidence to counter the earlier empirical studies which did not support the latter hypothesis. There are two reasons why our results could conflict with the earlier (pre-Chowdhury ) evidence. First, the years we consider are different. Second, these earlier studies did not use unit root and cointegration techniques in performing Granger causality tests.
Our results also conflict with those of Chowdhury . Again, one difference is the years; moreover Chowdhury considered the U.K. rather than the U.S., as we do. There are two other possible reasons why our results would differ from Chowdhury. First, a potentially serious problem is that he assumed a maximum lag length of four in his bivariate vector autoregressive (VAR) models, while our results suggest a longer lag length. Second, he used Granger causality tests based on levels VAR models. As Toda and Phillips  show, the causality test based on the error correction models, which we use in our analysis, performs better than the causality test based on levels VAR models.
Because this might suggest that policies to encourage advertising would increase aggregate demand and therefore could be used to control the business cycle, we should add a caveat. Increased consumption today means lower savings today. This in turn means lower investment. Certainly, more study is required before we could determine the total effect of advertising on the business cycle. Clearly, there is no reason to believe that advertising is a better candidate for countercyclical policy than other policies which have been tried, with all their attendant failures and unintended consequences, in the past.
1. Taylor and Weiserbs  seem unaware of previous empirical work. Regarding the controversy (which is usually attributed to Galbraith but, as Ekelund and Saurman  note, was actually initiated by Hansen ), they state that "no one [involved in] the debate [concerning the effect of advertising on aggregate consumption] has seen fit to examine by econometric methods the proposition that advertising has an impact on ... aggregate consumption ..." [26, 642]. They cite neither Verdon, McConnell, and Roesler nor Ekelund and Gramm.
2. In his 1972 book, Schmalensee tests each hypothesis using numerous functional forms, and uses the Houthakker-Taylor form (in addition to several others) in his examination of the aggregate advertising-aggregate consumption question.
3. The reader might wonder why we do not discuss a model developed by Berndt in his book. The reason is that Berndt did not develop one. While Berndt  examines the application of various econometric techniques to different economic questions and supplies excellent literature reviews, he does not engage in original research in this book.
4. Another study, which uses techniques similar to those we employ, is Baghestani  who examines the firm-level advertising-sales relationship. As we pointed out in the introduction, that issue is quite different from the aggregate advertising-aggregate consumption relationship.
5. We use the GNP deflator in order to increase our sample size to the years 1947-88.
6. For instance, compare the empirical estimates in Doroodian and Seldon  vs. those in Seldon and Boyd . While the objectives of the papers differ, both involve an estimation of cigarette demand. The former deflates advertising using the CPI, the latter deflates advertising using an advertising price index; and the results (quantitative and qualitative) are very close.
1. Ackley, Gardner. Macroeconomic Theory. New York: Macmillan, 1961.
2. Ashley, Richard, Clive W. J. Granger and Richard Schmalensee, "Advertising and Aggregate Consumption: An Analysis of Causality." Econometrica, July 1980, 1149-67.
3. Akaike, Hirotugu, "Statistical Predictor Identification." Annals of the Institute of Statistical Mathematics, Fall 1970, 203-17.
4. -----, "A New Look at the Statistical Model Identification." IEEE Transactions on Automatic Control, December 1974, 716-23.
5. Baghestani, Hamid, "Cointegration Analysis of The Advertising-Sales Relationship." Journal of Industrial Economics, December 1991, 671-81.
6. Berndt, Ernst R. The Practice of Econometrics: Classic and Contemporary. Reading, Mass.: Addison-Wesley, 1991.
7. Box, George E. P. and David Roxbee Cox, "An Analysis of Transformations." Journal of the Royal Statistical Society, Series B, 1964, 211-43.
8. Chowdhury, Abdur R., "Advertising Expenditures and the Macro-Economy: Some New Evidence." International Journal of Advertising, No. 1 1994, 1-14.
9. Clarke, Darral G., "Econometric Measurement of the Duration of the Advertising Effect on Sales." Journal of Marketing Research, November 1976, 345-57.
10. Dickey, David A. and Wayne A. Fuller, "Distribution of the Estimators for Autoregressive Time Series with a Unit Root." Journal of the American Statistical Association, June 1979, 423-31.
11. Doroodian, Khosrow, and Barry J. Seldon, "Advertising and Cigarette Consumption." Eastern Economic Journal, July-September 1991, 359-66.
12. Duesenberry, James S. Income, Saving, and the Theory of Consumer Behavior. Cambridge: Harvard University Press, 1967.
13. Duffy, Martyn, "Advertising in Demand Systems: Testing a Galbraithian Hypothesis." Applied Economics, March 1991, 485-96.
14. Ekelund, Robert B., Jr. and William P. Gramm, "A Reconsideration of Advertising Expenditures, Aggregate Demand, and Economic Stabilization." Quarterly Review of Economics and Business, Summer 1969, 71-77.
15. Ekelund, Robert B., Jr. and David S. Saurman. Advertising and the Market Process. San Francisco: Pacific Research Institute for Public Policy, 1988.
16. Engle, Robert F. and Clive W. J. Granger, "Co-integration and Error Correction: Representation, Estimation, and Testing." Econometrica, March 1987, 251-76.
17. Grabowski, Henry G., "The Effects of Advertising on Intraindustry Shifts in Demand." Explorations in Economic Research, Fall 1978, 675-701.
18. Granger, Clive W. J., "Some Properties of Time Series Data and Their Use in Econometric Model Specification." Journal of Econometrics, May 1981, 121-30.
19. Hansen, Alvin. Economic Issues of the 1960's. New York: McGraw-Hill, 1960.
20. Hay, Donald A. and Derek J. Morris. Industrial Economics: Theory and Evidence. Oxford: Oxford University Press, 1979.
21. Johansen, Soren, "Statistical Analysis of Cointegration Vectors." Journal of Economic Dynamics and Control, September 1988, 231-54.
22. ----- and Katarina Juselius, "The Full Information Likelihood Procedure for Inference on Cointegration - With Applications to the Demand for Money." Oxford Bulletin of Economics and Statistics, May 1990, 169-210.
23. MacKinnon, James G., Halbert White, and Russell Davidson, "Tests for Model Specification in the Presence of Alternative Hypotheses: Some Further Results." Journal of Econometrics, January 1983, 53-70.
24. Schmalensee, Richard. The Economics of Advertising. Amsterdam: North-Holland, 1972.
25. Seldon, Barry J. and Roy Boyd, "The Stability of Cigarette Demand." Applied Economics, February 1991, 319-26.
26. Taylor, Lester D. and Daniel Weiserbs, "Advertising and the Aggregate Consumption Function." The American Economic Review, September 1972, 642-55.
27. Toda, Hiro Y. and Peter C. B. Phillips, "Vector Autoregressions and Causality." Econometrica, November 1993, 1367-93.
28. Verdon, Walter A., Campbell R. McConnell, and Theodore W. Roesler, "Advertising Expenditures as an Economic Stabilizer: 1945-64." Quarterly Review of Economics and Business, Spring 1968, 7-18.
|Printer friendly Cite/link Email Feedback|
|Author:||Seldon, Barry J.|
|Publication:||Southern Economic Journal|
|Date:||Jan 1, 1995|
|Previous Article:||The Business Cycle: Theories and Evidence.|
|Next Article:||Nominal and real wage cyclicality during the interwar period.|
|Broadcast advertising and U.S. demand for alcoholic beverages.|
|The Dynamic Behavior of Wages and Prices: Cointegration Tests within a Large Macroeconomic System.|