Sie sind auf Seite 1von 28

International Journal of Industrial Organization 18 (2000) 11531180 www.elsevier.

com / locate / econbase

Media substitution and economies of scale in advertising


Barry J. Seldon a , *, R. Todd Jewell b , Daniel M. OBrien a
a

School of Social Sciences, University of Texas at Dallas, Richardson, TX 75083 -0688, USA b Department of Economics, University of North Texas, Denton, TX 76203 -3677, USA

Received 1 January 1998; received in revised form 1 December 1998; accepted 1 March 1999

Abstract Two important issues in the economics of advertising are media substitutability in generating sales and scale economies in advertising. If media are substitutes then partial bans, e.g., broadcast bans on cigarettes or alcoholic beverages, may be ineffective; and mergers among radio and TV rms, currently widespread in the U.S. and Mexico, are unlikely to result in market power in setting advertising rates. If there are scale diseconomies in advertising, concerns that advertising increases entry barriers may be unfounded. Using U.S. beer rm data over 1983:Q11993:Q4 for three media categories, we nd evidence of high substitutability and diseconomies of scale. 2000 Elsevier Science B.V. All rights reserved.
Keywords: Advertising; Economies of scale; Media substitution; Advertising bans; Beer JEL classication: L13; M37; D21

1. Introduction Two signicant issues in the economics of advertising involve the substitutability of advertising media and the existence of scale economies in advertising.1 These
* Corresponding author. Tel: 11-972-883-2043; fax: 11-972-883-2735. E-mail address: seldon@utdallas.edu (B.J. Seldon). 1 We are using the terminology scale economies in advertising in the usual, but admittedly imprecise, way. As Waterson (1984, p. 134 and the associated footnote 8 on p. 219) states, This is a very common terminology but is obviously slightly inaccurate; what we really mean is a range of output over which there are increasing returns to advertising expenditures. We discuss the concept further below. 0167-7187 / 00 / $ see front matter 2000 Elsevier Science B.V. All rights reserved. PII: S0167-7187( 99 )00010-7

1154

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

issues are important in understanding both rm behavior and the relative impact of the various advertising media. In addition, they are relevant to three separate public policy debates: the effectiveness of partial advertising bans; the effects of the current merger wave among radio and television rms; and the anticompetitive effects of advertising. In this paper, we explore these issues using a translog cost model which estimates the cost of advertising in various media in order to sell particular quantities of a good. We use quarterly data for U.S. beer rms over 1983 to 1993 and separate advertising into three media categories: print, television, and radio. While we analyze a particular market, our qualitative results may apply to many other markets. The beer industry is representative of industries which produce alcohol beverages; and the consumption of these beverages is central to current policy debates. Various public policies have been designed to reduce the negative effects of alcohol use, which include cirrhosis of the liver, drunk-driving fatalities, and alcohol-related crime. Policies such as alcohol taxation, minimum drinking age laws, and alcohol availability restrictions are designed to decrease the consumption of alcohol by increasing its price or by limiting access to alcohol beverages (Grossman et al., 1993; Wagenaar, 1993; Jewell and Brown, 1995). Another tack pursued to lower the consumption of alcohol involves advertising. For some time, public policies have been proposed which would either ban beer and wine advertising in the broadcast media (similar to the ban on cigarette advertising, see, e.g., Fritschler, 1989) or restrict alcohol advertising in all media, because many people believe that advertising increases market demand.2 Recent events reignited the alcohol advertising controversy in the U.S. In May 1996, Representative Joseph Kennedy II introduced legislation addressing the advertising of alcohol beverages. His proposal would require health warnings, eliminate tax deductions for advertising, limit print advertising to black and white text in many publications, and restrict most TV advertising to the hours between 10:00 p.m. and 7:00 a.m. A month later, despite this attack on alcohol advertising, Seagram initiated TV advertising in Texas. This broke with the tradition among

2 This contrasts with the alcohol industries claim that advertising merely redistributes market share and has no effect on total alcohol consumption. The connection between alcohol consumption and advertising has been investigated, but the results of these studies are mixed (for literature reviews, see Saffer, 1993; Smart, 1988). Smart (1988) reported that earlier studies found little effect of advertising on alcohol consumption and that advertising bans have little effect on alcohol sales. But the controversy continued in later publications. Saffer (1991) concluded that banning alcohol advertising would reduce consumption. Young (1993), however, reexamined Saffers data and found that advertising bans may in fact lead to increased alcohol consumption. On the other hand, Tremblay and Tremblay (1995) found that a beer rms advertising increases the rmss inverse demand function and that there are positive spill-overs from rivals advertising. Because this is true for all beer rms, these results imply that advertising increases the market demand for beer; which in turn suggests that bans could reduce consumption. Lee and Tremblay (1992) suggested that, while advertising bans may reduce consumption, tax increases could be more effective.

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1155

liquor producers of avoiding radio (and, later, TV) advertising since 1936, a tradition formalized in 1948 by the Distilled Spirits Councils (DSCs) Code of Good Practice. Representative Kennedy responded by introducing another bill to ban liquor ads on TV and radio (Beatty, 1996a,b,c). Despite this attack, the DSC voted unanimously to follow Seagram in overturning the voluntary ban on TV and radio advertising (Beatty, 1996e). This provoked President Clinton to urge the Federal Communications Commission (FCC) to ban liquor advertising on TV (Associated Press, 1996). The Federal Trade Commission (FTC) acted rst by opening an investigation of alcohol beverage advertising (including beer), while the FCC contemplated, but to date has not undertaken, a parallel investigation (Ingersoll, 1996). Instead, the FCC suggests that broadcasters refuse liquor ads (Pope, 1997). Most recently, the FTC forced two TV commercials, for Becks beer and premixed Kahlua White Russian cocktails, off the air (Beatty, 1998). However, the effectiveness of bans on TV and radio advertising remains an open question. It depends not only upon whether such bans reduce market consumption (as suggested for alcohol by Saffer (1991), and implied by the results of Tremblay and Tremblay (1995, 1997)), but also upon whether other media are close substitutes for broadcast advertising. If they are close substitutes, the partial ban on advertising proposed by Representative Kennedy and President Clinton may not be effective. In this paper, we consider the substitutability of advertising media by beer rms. While the results are important to a consideration of partial advertising bans on beer, they may also be suggestive for advertising bans on liquor, tobacco, or other products. If media are substitutable in the beer industry, we might expect media substitutability in other markets. There is yet another aspect of media substitutability that is important for public policy. With recent mergers among entertainment corporations, especially since the passage of the Telecommunications Act of 1996, concerns have been raised that market power might accrue to the owners of newly-merged television or radio stations that sell advertising spots (Beatty, 1996d). This, in fact, is the subject of ongoing research at the U.S. Department of Justice and the Mexican Federal Competition Commission (Comision Federal de Competencia) and is an implication of recent research concerning substitutability at the FCC. If media substitutability is similar across markets, the results of this study will suggest whether concerns about market power are well grounded or if competitive pricing for advertising outlets are ensured by easy substitution into other media.3 While we use U.S. data in this study, the results may be generalizable to other countries.

Research at the FCC and the Mexican CFC has been discussed with one of the authors of this study in personal communications with economists at both Commissions. However, the current study is independent of their efforts and is more encompassing. For instance, FCC research considered substitutability between different types of TV advertising only (McCullough and Waldron, 1998).

1156

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

To address these issues, we apply the little-used, but appropriate, Berndt and Savin (1975) autocorrelation correction to a translog cost function. We explore media substitution by estimating Morishima elasticities, which Blackorby and Russell (1989) demonstrated to be preferable to the more conventional Allen elasticities. We uncover evidence suggesting that there are diseconomies of scale in advertising. In addition, we nd the various media to be substitutable.

2. Media substitutability, advertising economies, and public policy We will estimate advertising economies of scale and substitution elasticities among advertising media using an advertising cost function, which is the (minimum) total cost of advertising when the rm uses several media to generate a particular level of sales. Assuming separability of the production function, which relates levels of production inputs to the level of output, and the advertising function, which relates the number of advertising messages in various media to the level of sales, we can employ Browns (1978) denition of scale economies in advertising (which is consistent with Waterson, 1984) as a greater than proportional increase in quantity sold per given increase in units of advertising. This denition is common in the advertising literature, so our results will be conceptually comparable to previous results. Our substitution elasticities will indicate how media can be substituted to maintain sales. While this paper differs from most previous studies, two articles explore these issues in manners comparable to ours. Bresnahan (1984) estimated advertising cost models for two cases: a cross-industry study of six consumer goods and a single-industry study of beer producers. While he separated advertising by media, Bresnahan was concerned with economies of scale in advertising and substitution between retail services and advertising media. He did not address substitution among the media. Seldon and Jung (1993) considered the markets for advertising in four categories of media (print, radio, TV, and outdoor), in effect summing across all products in the U.S. economy. They did not estimate economies of scale. They found the different media to be fairly good substitutes. However, because they summed advertising in each medium across industries and because all industries do not advertise in all media, their results should only be taken as suggestive, as they acknowledged. This study differs from these previous efforts in focus, modeling, and data. While Bresnahan estimated only share equations for the beer industry, we simultaneously estimate a cost function and share equations and focus upon substitution elasticities among the various media. We also consider economies of scale, but in a manner which differs from Bresnahans approach; we discuss this below. In addition, Bresnahans estimates used annual data from 1967 to 1981, while our study considers more recent quarterly data. Our study differs from Seldon and Jung in two respects. First, we concentrate on the rm level in one industry. This affords insight concerning the substitutability of media in the sales

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1157

of a particular good. It also permits a discussion of policy issues related to advertising in that industry. Second we use quarterly, instead of annual, data.4 Our use of quarterly advertising data, which distinguishes our study from most previous studies of advertising, is more critical than one might think. The use of annual advertising data when advertising effects upon demand depreciates in a shorter period may result in a temporal aggregation bias which seriously distorts estimated coefcients (Bass and Leone, 1983; Leone, 1995). And there is strong evidence that a year is too long a period. Boyd and Seldon (1990), Thomas (1989), and Seldon and Doroodian (1989), who used annual data, suggest that advertising effects depreciate within a year; while Ashley et al. (1980) and Leone (1995), who examined shorter periods, suggest that advertising effects may depreciate within a quarter. In fact, Leone points to evidence that advertising effects for frequently purchased grocery products are largely depreciated within six weeks. Because beer is frequently purchased, we use quarterly, rather than annual, data. We investigate advertising in the beer industry over the quarters 1983:Q1 1993:Q4. Advertising messages are divided into three media categories: those which are visual (print); those which are audio (radio); and those which are audiovisual (TV). The print media include magazines, newspapers, and outdoor advertising. We estimate a rm-level translog advertising cost function associated with a given level of sales, with media share equations as side conditions. The results will indicate a potential response by rms to a partial advertising ban in the beer market. The results also will suggest whether the mergers of TV and radio companies threaten the competitive pricing of advertising. In addition, the results will provide information on economies of scale in advertising. If there are economies of scale in advertising and if the maintenance of high levels of advertising by market incumbents is credible in the face of entry, advertising scale economies may raise barriers to entry because entrants would either have to incur a higher unit cost for generating sales or enter the market at a larger advertising scale.5 Under these conditions, policies that discourage advertisThere are still other studies of the beer industry that estimate substitution elasticities from a cost function which does not separate production and advertising costs, and where advertising is not separated by media (e.g., Tremblay, 1987). Our approach is similar to Bresnahan (1984), who noted that if the production cost and advertising cost functions are separable (which is perfectly reasonable) then one may consider the advertising cost function in isolation. This model is developed and further justied below. 5 We address advertising economies of scale, but there are several means through which advertising could conceivably be anticompetitive, including an incumbent investing in long-lived advertising. However, as discussed previously, recent research found that advertising depreciates rapidly, so this is unlikely. Advertising may also have procompetitive effects, such as allowing an entrant to advertise before entering a market. As an example of this, Gallo entered the wine cooler market with advance advertising for Bartles and Jaymes. For discussions of possible anticompetitive and procompetitive effects of advertising see Albion and Farris (1981), Ekelund and Saurman (1988), Hay and Morris (1991, pp. 142148), Krouse (1990, pp. 498501), McAuliffe (1987), Scherer and Ross (1990, pp. 406407), Sutton (1991), and Waterson (1984, pp. 134136 and 201204).
4

1158

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

ing may encourage lower prices due to the threat of entry if higher prices are maintained. In addition, previous research (see, e.g., Eckard, 1991) suggests that, if advertising is persuasive rather than informative, higher levels of advertising are associated with lower levels of price competition. Thus, in markets with persuasive advertising, policies that discourage advertising may further encourage price competition.

3. The cost function We employ duality techniques most often applied to the production of goods and services. We assume that the rm produces and sells beer using different inputs for production than for advertising. A production function relates the levels of inputs to the level of output, while our advertising function relates advertising messages to sales. Advertising enables sales for at least one of two reasons. First, advertising informs consumers that the good, a brand of beer in our case, exists. Second, advertising may serve in a persuasive capacity, encouraging the consumer to try the brand. Because different inputs are used for production and advertising, the cost function for production and advertising is separable in the sense that we can specify a production cum advertising cost function # (W, P; Q) 5 G(W; Q) 1 C(P; Q) where W is a vector of prices of production inputs, P is a vector of prices of advertising messages in various media, G is the production cost function given that the rm wishes to produce Q units, and C is the advertising cost function given that the rm wishes to sell Q units. This allows us to concentrate on the rms advertising cost function, C(P; Q), which gives the cost of advertising required to sell Q units. An alternative reason for separating the advertising cost function from the production cost function was suggested by Seldon and Jung (1993). Production occurs prior to advertising in the sense that rms may plan their production and advertising simultaneously; however, once production has occurred, rms are free to reconsider advertising plans made during production. For instance, if inventories were accumulating unexpectedly, rms could increase the number of advertisements (e.g., by renting more billboard space) relative to anticipated levels.6 The prices of advertising messages in the P vector are exogenous to the model. These prices are assumed to be xed by the markets for advertising messages, and beer advertising is a small fraction of all advertising in every medium. As Bresnahan (1984, pp. 139140) pointed out, when we consider the costs of production and advertising as being separable, it makes sense to discuss economies

In this case, the quantity that the rm wishes to sell can be treated as a predetermined variable in the econometric estimation. We test this assumption below.

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1159

of scale in advertising. That is, we can consider what happens to sales when we increase all advertising in the various media proportionately, as in Brown (1978), and not be concerned with simultaneously increasing all production inputs by the same proportion.7 The advertising cost minimization problem may be stated as: min C 5 AP such that Q 5 C (A)
A

where C is the total cost of advertising, A is a row vector of advertising messages in various media, P is a column vector of the prices of messages in the media, and Q is the output that rms wish to sell.8 C (A) is a quasiconcave twice differentiable advertising function that relates the quantity that the rm wishes to sell to the number of advertising messages used in each medium. The minimization problem results in the rms cost function C 5 C(P; Q). We employ a exible functional form, the translog model 9 ln Cf,t 5

a0 1 fq (ln Q f,t ) 1 (1 / 2)fqq (ln Q f,t )2


1
i i,t ij i i j

O a (ln P ) 1 (1 / 2) O O b (ln P )(ln P ) 1O g (ln Q )(ln P ) 1 u


i,t j,t qi f,t i,t f,c,t i

(1)

where Pi,t is the price of advertising messages for medium i, j 5 p, v, r ( p5print, v5television, and r5radio); f is the rm index; t is the time index; and uf,c,t is an error term. By the symmetry of second-order coefcients, bij 5 bji By the form of Eq. (1) we assume, as did Caves et al. (1981), that parameters

This, of course, is a reversal of the usual economies of scale study which (1) ignores advertising and (2) considers what happens to output when all production inputs increase proportionately. For instance, Elzinga (1995, pp. 129131) implicitly ignores advertising costs in his discussion of economies of scale in beer production while later in the article he discusses the marketing of beer. Even Sutton (1991), who considers advertising in great detail throughout much of the book, discusses economies of scale in production (in Appendix 13, entitled Scale Economies in Brewing) as separate from advertising. When Sutton considers the rising trend in concentration in recent years in the U.S. market, he notes (p. 299) that it can . . . be argued that the initial impetus [was] the changing degree of scale economies [in production] in the industry, but also that a central role . . . was played by the . . . escalation of advertising outlays. Still, Sutton does not directly address economies of scale in advertising, as we do in this paper. 8 The advertising function might be written more completely as C (A, A Q ) where A Q is the advertising of the rms rivals. We do not have data for all beer producers, so we are unable to construct A Q . As another alternative, we might write the function as C (A; 3 ) where 3 is a vector of prices of the goods being sold (or, alternatively, an average price for the rms products). We are unable to nd rm-level price data. We explain how we treat this potential missing data problem below. A Hausman test supports our treatment. 9 Translog models are reviewed in Berndt (1991, pp. 469487).

1160

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

are the same across rms. This is plausible in the case of advertising cost so long as all rms have access to the same advertising media, which they do, and so long as the inputs used to create the ads are hired in competitive markets, as they seem to be (Jung and Seldon, 1995). In fact, after obtaining the results reported below, we added rm dummy variables to the intercept of Eq. (1), as in Bresnahan (1984), but found little change in our results. The difference among rms seems to be reected in size: a regression of ln Q f,t on rm dummies explained 95 percent of the variance.10 Therefore, the addition of rm dummies would induce multicollinearity. Moreover, if rms have access to the same media and create ads through competitive markets, including the dummies would cause the model to be misspecied. Taking the derivative of Eq. (1) with respect to ln Pi,t and applying Shephards Lemma yields the cost share equations for the media. They are: Sf,i,t 5 ai 1

O b (ln P ) 1 g (ln Q
ij j,t qi j

f,t

) 1 uf,i,t

(2)

for i, j 5 p, v, r where Sf,i,t 5 Pi,t A f,i,t /Cf,t ; A f,i,t is the number of messages in the ith medium; and uf,i,t is an error term. The cost function must be homogeneous of degree one in prices. This homogeneity condition requires the following restrictions:

O a 51
i i

and

O b 5O b 5O g 5 0.
ij ij qi i j i

(3)

4. The data Our data cover the period 1983:Q1 through 1993:Q4, which contains 44 quarters. Monetary variables are deated to 1982 dollars using the producer price index from the Economic Report of the President. The variables, their sources, and means and standard deviations for the different rms are presented in Table 1. Beer advertising data are from various issues of Competitive Media Reportings Ad $ Summary (Competitive Media Reporting, 198393). This publication reports

10

The estimated regression equation is ln Q f


2

5 1.24 (0.03)

1 1.71DA (0.04)

1 0.19DC (0.04)

2 1.50DG (0.07)

2 0.16DP (0.10)

1 0.31DS (0.05)

with R 5 0.95 where Dx is a dummy variable for rm x 5 A (Anheuser-Busch), C (Coors), G (Genessee), P (Pabst), and S (Strohs). The omitted rm is Heilman / Bond. Standard errors are in parentheses. In this regression, the intercept and four of ve dummy variables are signicant at better than the 1 percent level. The dummy for Pabst is signicant at the 10 percent level.

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180 Table 1 The variables, their sources, and descriptive statistics Quarterly advertising expenditures and sales Variable, source, units Total advertising expenditures Source: Ad $ Summary Units: 1000s, 1982 dollars Advertising expenditures Television Source: Ad $ Summary Units: 1000s, 1982 dollars Advertising expenditures Print Source: Ad $ Summary Units: 1000s, 1982 dollars Advertising expenditures Radio Source: Ad $ Summary Units: 1000s, 1982 dollars Beer sales Source: Beverage Industry Units: millions of barrels Firm Anheuser-Busch Coors Genessee Heilman / Bond Pabst Strohs Anheuser-Busch Coors Genessee Heilman / Bond Pabst Strohs Anheuser-Busch Coors Genessee Heilman / Bond Pabst Strohs Anheuser-Busch Coors Genessee Heilman / Bond Pabst Strohs Anheuser-Busch Coors Genessee Heilman / Bond Pabst Strohs Medium Television Print Radio Mean 69 852.32 20 429.18 1312.63 4511.23 3550.59 9477.01 61 209.88 18 827.22 1192.96 3636.20 3355.34 8515.26 5402.70 1240.14 119.67 662.46 195.25 666.89 3239.73 361.81 0.00 213.11 0.00 294.86 19.34 4.22 0.77 3.53 2.95 4.85 Mean 1.207 1.195 1.107 Standard deviation 14 396.35 7087.35 461.77 2851.42 1655.53 6123.33 12 401.69 6233.48 437.92 2645.00 1637.86 5978. 70 2046. 96 884. 50 72.25 574.54 111.21 601.03 2959.84 949.28 0.00 682.65 0.00 558.87 2.77 0.69 0.09 0.76 0.37 1.19 Standard deviation 0.104 0.091 0.067

1161

Price indexes Source and units Source: McCann-Erickson, Inc. Units: 1982-based Index

nominal advertising expenditures in the various media for the 1000 largest advertisers in the U.S. Thus, the beer rms in our data set are companies that sell national or large regional brands; namely Anheuser Busch, Coors, Genessee,

1162

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

Heilman / Bond, Pabst, and Strohs.11,12 In some quarters, data for some rms are not reported because their quarterly advertising expenditures fell sufciently so they were not among the 1000 largest advertisers. Nevertheless, we collected 195 observations. Deating advertising expenditures in the various media to 1982 dollars, we obtain Pi,t A f,i,t of Eq. (2). The Ad $ Summary also reports rms total advertising expenditures across media. Deating these gures yields Cf,t of Eq. (1). In a few cases, the sum of a rms media expenditures does not equal a rms reported total expenditures. In these cases, we replace the reported total with the sums of the reported media expenditures because the reported total expenditures seemed to be out of line with the time series for the particular rms. McCann-Erickson, a New York City advertising agency, provided cost indexes for advertising messages, which we employ for prices (Pi,t ).13 We use their cost-per-thousand media indexes, which represent the media advertising cost for reaching a thousand-person audience. The cost indexes are based on 1982 prices and are constructed by dividing the total cost of advertising in the various media by the number of people (in thousands) exposed to the advertising. Hence, advertising messages are conceptually measured in units which reach 1000 people. From the less aggregated McCann-Erickson cost indexes, we calculate Divisia indexes for the costs of messages in print, television, and radio using the ratio of expenditures in particular media (summed across the rms) to aggregated advertising expenditures (summed across rms and media) as weights. Thus, all rms face the same prices for advertising in a given medium. For example, we calculate a cost index for print media from data for magazines, newspapers, and outdoor advertising. This captures all print media in which beer rms advertise.14 While an element in any of these index series is the nominal cost-per-thousand in year t divided by the nominal cost in 1982, the indexes still increase with ination. To make these data compatible with the other monetary data, we deate the indexes by the 1982-based producer price index. The resulting advertising price indexes for the different media are annual series. We create a quarterly series for each medium using cubic spline functions (Greene, 1993). This method joins cubic polynomials together to form a continuous time series with continuous rst

11 The Beverage Producers Association reports that the rms in this list accounted for 73.6 percent of domestic beer sales in the U.S. in 1993. We exclude Miller, a subsidiary of Philip Morris, because we are unable to isolate beer advertising expenditures. We exclude imported beer because a reliable time series for sales is not available for the entire 11-year period. 12 Because our sample includes only large beer producers, our results may not be accurate for smaller rms. Data for smaller rms are not available. 13 McCann-Erickson Inc. supplies the advertising data that are published annually in the Statistical Abstract of the United States. The price data are available from the authors upon request. 14 Beer is not advertised through direct mail, a medium used for many other goods and services.

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1163

and second derivatives. The method assures that the average of the created quarterly data in each year equals the observed price index for that year.15 To estimate the cost function, we need rm-level sales data (Q f,t ). Data for the quantity of beer (millions of barrels) sold by each rm in each year are available from various issues of Beverage Industry. These quantities are annual data and must be converted to quarterly data. Because advertising expenditures exhibit seasonality, we convert the sales data in a manner which preserves the seasonality of sales. The Beer Institute (1994) records total U.S. beer sales on a monthly basis over the years of our sample. We convert this series to a quarterly series by adding sales for each month of each quarter and then determine the percentage of annual sales (in 1982 dollars) for each quarter for every year of our sample. We then apply these seasonal percentages to the various rms annual sales to compute a quarterly sales series for each rm.

5. The econometric model To simplify notation, we hereafter drop the rm index f. We estimate Eqs. (1) and (2) subject to conditions (3) using Zellners iterated seemingly unrelated regressions (ITSUR). To avoid singularity of the covariance matrix, one share equation must be omitted. ITSUR produces coefcient estimates which are invariant with respect to the omitted equation (Zellner, 1962, 1963). We omit the radio advertising share equation, but we recover the parameters of this equation using restrictions given by Eqs. (3). We also estimate additional sets of equations to test certain restrictions on the cost function. These tests will suggest the appropriate functional form of the regression equations to use in further analysis. One restriction (gqi 5 0 for i 5 p, v, r) forces the advertising function C (A) to be homothetic, implying that the advertising cost function is separable in output and prices (Christensen and Greene, 1976). If this condition holds, we could write the cost function as C 5 C1 (Q)C2 (P). A second restriction forces the advertising function to be homogeneous. The homogeneity restriction is that gqi 5 0 (as before, since any homogeneous function is homothetic) and gqq 5 0, so that the elasticity of cost with respect to output is constant. If the advertising function is homogeneous and fq 5 1, there are constant returns to scale in advertising. Because advertising effects depreciate rapidly, probably within months, we assume that any effects of previous quarters advertising upon present sales are small enough to disregard. Thus, if any effect of past advertising actually exists, it

The software used to produce the quarterly series is the EXPAND procedure of SAS Version 6. For a more complete explanation, see SAS Institute (1988, pp. 261277).

15

1164

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

is suppressed into the error terms of the model.16 This would create patterns in the time series of the error terms and can be treated, and corrected for, as autocorrelation using the method of Berndt and Savin (1975).17 Care must be taken when correcting for autocorrelation in the share equations. First, the lagged error term associated with one equation may affect other equations (see, e.g., Berndt, 1991, pp. 476479). Second, we must ensure that the shares sum to one. Write the system of share equations as S t 5 G X t 1 U t ; t 5 2, 3 . . . , 44 where S t is the three-element matrix of media shares (including the share which will be excluded in the econometric estimation), G is the matrix of parameters of the system suggested by Eq. (2), X t is the vector of predetermined variables, and U t is the vector of error terms associated with the share equations. Now, in addition to ensuring homogeneity of the cost function, conditions (3) ensures that i G Xt 5 1, where i 5 (1 1 1). Therefore, in order to obtain the adding-up condition, all that remains is to ensure that u r,t 5 2 u p,t 2 u v,t . Let ri, j be the autocorrelation coefcient associated with the lagged error term of the share equation of the jth advertising medium as it affects the ith share equation. Following Berndt and Savin (1975), we specify our three-element vector U t as:

343

u p,t rp, p 2 rp,r rp,v 2 rp,r u v,t 5 rv, p 2 rv,r rv,v 2 rv,r u r,t rr, p 2 rr,r rr,v 2 rr,r

4F

np,t u p,t21 n u v,t 21 1 v,t nr,t

34

where ni,t for i 5 p, v, r are well-behaved error terms. This may be done because, in order to assure that shares sum to one, each column of the original undifferenced autocorrelation matrix with element ri, j (where i, j 5 p, v, r) must sum to the same, but unknown, constant number. We cannot solve for ri, j , but there is no need to do so. The above discussion does not take into consideration the system of equations which includes the cost function, but the extension is straightforward. As Berndt and Savin (1975, p. 955) pointed out, the restrictions discussed above apply only to autocorrelation terms associated with share equations. When we include the cost function, each of the share equations includes the lagged error of the cost function, and the cost function includes lagged errors for the share equations taking into

16 Because of missing quarterly observations as discussed in the data section, any attempt to incorporate lagged advertising would reduce our sample size considerably. 17 Similarly, if the advertising function were more properly specied as in footnote 8, the effects of any missing variables will also be relegated to the error terms, requiring us to treat the error terms for autocorrelation.

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1165

consideration the adding-up constraint. The error terms that are used in the econometric estimation are, then, u c,t 5 rc,c u c,t 21 1 ( rc, p 2 rc,r )u p,t 21 1 ( rc,v 2 rc,r )u v,t 21 1 nc,t u p,t 5 rp,c u c,t 21 1 ( rp, p 2 rp,r )u p,t 21 1 ( rp,v 2 rp,r )u v,t 21 1 np,t and u v,t 5 rv,c u c,t21 1 ( rv, p 2 rv,r )u p,t 21 1 ( rv, p 2 rv,v )u v,t 21 1 nv,t (6) (4) (5)

where the terms for the r -differences (viz., ri, j 2 ri,k ; i, j, k 5 c, p, v, r) are estimated as one parameter because we are not interested in, and cannot estimate, these r s individually. Eqs. (4), (5), and (6) are substituted directly into the cost Eq. (1) and the share Eqs. (2), and the system is estimated using nonlinear iterated seemingly unrelated regressions (NLITSUR). This nonlinear method allows us to estimate the model parameters (a s, b s, g s, and f s) and the autocorrelation parameters ( ri,c , i 5 c, p, v; and the r -differences) simultaneously. This is because, in the place of the lagged error terms in Eqs. (4)(6), we substitute the difference between the lagged dependent variable and the right-hand-side function with lagged independent variables. Thus, the lagged error term for a share equation is u i,t 21 5 Si,t 21 2 ai 1

F O b (ln P
ij j

j,t 21

) 1 gqi (ln Q t 21 )

for i, j 5 p, v, r. The lagged error of the cost function is estimated similarly. This is superior to other methods, such as ITSUR combined with the iterated CochraneOrcutt autocorrelation correction, which can only approximate the more exact estimates we are able to obtain using NLITSUR. The autocorrelation coefcients are then subject to the same asymptotic t tests applied to the model parameters.18 Given the structure of Eqs. (4)(6), we are able to use only observations for which lagged data are available. Thus, we lose some observations due to missing data for some rms for some quarters. Nevertheless, we retain a sample size of 177.

6. Estimation of the cost and share functions The results of the NLITSUR estimations for the system of equations with

Because the inclusion of a lagged error term necessitates inclusion of the lagged endogenous variable, it might seem that we could distinguish between short- and long-run elasticities as in partial adjustment models. However, the inclusion of a lagged error term does not impart any dynamics to the economic model. See Seldon and Bullard (1992) for clarication.

18

1166

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

autocorrelation (obtained by substituting Eqs. (4)(6) into the cost and share equations given by (1) and (2)), and the systems of equations for the autocorrelated homothetic, homogeneous, and constant returns to scale (constant RTS) models are presented in Table 2.19 The signicance of the majority of the autocorrelation coefcients indicate that our use of the Berndt-Savin correction is appropriate. All of the separate r s (v iz. rc,c , rp,c , and rv,c ) fall within the interval (21, 1), as we would expect, and the differenced r terms are plausible. For example, the estimates for ( rc,v 2 rc,r ), which are greater than one, could occur if rc,v were in the interval (0, 1) and rc,r were negative and of sufcient magnitude within the interval (21, 0). Before further considering the results reported in Table 2, we test the propriety of treating the quantity Q as exogenous rather than endogenous using a J test (Davidson and MacKinnon, 1981, 1993) applied to the general cost function. Denote our cost equation as ln Ct 5 g(X g,t , b ) 1 eg,t where the exogenous quantities are included with the other right-hand-side variables in the X g vector, the b vector includes the model and autocorrelation coefcients, and eg is an error term. This model yields the estimated cost function g ; g(X g,t , b g ). Denote an alternative cost function as ln Ct 5 h(X h,t , b ) 1 eh,t where the vector X h is the same as X g with the crucial exception that, in every instance, the exogenous quantity variable, ln Q, is replaced with an instrumental variable created by regressing ln Q on the other exogenous variables in the system as well as on rm dummies and real disposable income.20 This model yields the estimated cost function h ; h(X h,t , bh ). The J test then involves running the two regression equations ln Ct 5 (1 2 ah )g(X g,t , b ) 1 ah h 1 egg,t and ln Ct 5 (1 2 ag )h(X h,t , b ) 1 ag g 1 ehh,t separately. If ah is signicant and ag is insignicant, then the model which uses the quantity instrument is the correct specication. Alternatively, if ah is insignicant while ag is signicant, then the model which uses the exogenous

19 We used Newtons method as the maximization procedure. The results reported in Table 2 appear to be the global maxima; when we initiated the coefcients at widely different values they converged to the same point. 20 A price variable for the nal product is not included as a determinant of ln Q in this test under assumption that the products price is endogenous. However, we ran another J test where we included a market-level beer price index (rm-level prices are not available), and our results in that test were qualitatively the same as our results for the J test described below.

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1167

Table 2 NLITSUR estimates of total cost and share function coefcients in the advertising cost model, 19831993 a
Coefcients General model 6.8361 d (1.5237) 0.1100 (0.2843) 0.8540 d (0.3302) 0.0361 (0.3690) 20.2997 c (0.1487) 20.1715 c (0.0748) 0.4713 d (0.1232) 0.1984 b (0.1417) 20.0269 (0.1481) 20.4443 d (0.1279) 0.1337 d (0.0294) 20.1417 d (0.0369) 0.0080 (0.0739) 0.8549 d (0.3046) 20.0945 (0.1341) 0.6068 d (0.0603) 20.5584 (0.5980) 1.2217 d (0.4944) 20.0227 d (0.0071) 0.9111 d (0.0670) 0.0472 (0.0590) 0.0366 d (0.0116) 20.6826 d (0.1234) 0.2277 d (0.0949) Homothetic model 6.8776 d (0.2484) 0.1240 d (0.0206) 0.8490 d (0.0279) 0.0270 (0.1341) 20.0953 (0.1275) 20.1008 (0.0823) 0.1961 c (0.1078) 20.0582 (0.1334) 0.1590 (0.1477) 20.3551 d (0.1080) 1.3704 d (0.2227) 20.0639 (0.1228) 0.6144 d (0.0612) 20.3951 (0.6511) 1.0022 c (0.4791) 20.0299 d (0.0073) 0.7042 d (0.0781) 0.1154 c (0.0578) 0.0438 d (0.0117) 20.6086 d (0.1269) 0.1602 c (0.0928) Homogeneous model 6.9518 d (0.2068) 0.1236 d (0.0206) 0.8498 d (0.0279) 0.0266 c (0.0145) 20.0888 (0.1268) 20.1009 (0.0821) 0.1896 c (0.1063) 20.0675 (0.1326) 0.1684 b (0.1068) 20.3580 d (0.1076) 1.2589 d (0.0735) 0.6113 d (0.0611) 20.4066 (0.6518) 1.0125 c (0.4790) 20.0307 d (0.0073) 0.7026 d (0.0778) 0.1184 c (0.0576) 0.0447 d (0.0118) 20.6081 d (0.1266) 0.1561 c (0.0926) Constant RTS model 16.8747 d (5.4142) 20.6861 b (0.4537) 2.1100 d (0.6894) 20.4239 b (0.2656) 20.0492 (0.1220) 20.0993 (0.0809) 0.1485 b (0.1005) 20.0980 (0.1288) 0.1973 c (0.1048) 20.3457 d (0.1082) set equal to 1.0 0.7697 d (0.0523) 20.5792 (0.6953) 1.4123 d (0.4886) 20.0394 d (0.0063) 0.6858 d (0.0781) 0.0967 c (0.0557) 0.0561 d (0.0106) 20.5979 d (0.1275) 0.1922 c (0.0909)

a0 ap av ar bpp bpv bpr bvv bvr brr gqp gqv gqr fq fqq rc,c
( rc, p 2 rc,r ) ( rc,v 2 rc,r )

rp,c
( rp, p 2 rp,r ) ( rp,v 2 rp,r )

rv,c
( rv, p 2 r,r ) ( rv, p 2 rv,v )

Goodness of t e 0.9984 0.9985 0.9985 0.9983 Degrees f of freedom 170.667 171.333 171.667 172.000 a Standard errors are in parentheses. The coefcient subscripts are p5print, v5television, r5radio, q5quantity to be sold, c5total cost. b Signicant at the 10 percent level, one-tailed test. c Signicant at the 5 percent level, one-tailed test. d Signicant at the 1 percent level, one-tailed test. e The goodness-of-t measure for the NLITSUR systems is McElroys (1977) R 2 . z f Degrees of freedom are equal for all equations in each of the systems. Because parameters are shared across equations, degrees of freedom are allocated across equations and need not be integers. Typically, the degrees of freedom for the cost equations would be different from the degrees of freedom for the share equations; but in our case they happen to be equal.

1168

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

quantities is the correct specication.21 In our case, ah is estimated as 20.25 with a t ratio of 20.64 while ag is estimated as 1.04 with a t ratio of 12.07. Therefore, our treatment of quantities as exogenous is appropriate. We now return to the models of Table 2. We test the validity of the homothetic, homogeneous, and constant RTS models using the likelihood ratio statistic 2 2(ln l) 5 nhln[det(Vr ) / det(Vu )]j where det(Vr ) and det(Vu ) are determinants of the estimated n -error covariance matrices of the restricted and unrestricted systems. The null hypotheses (that the restrictions hold) are subject to a x 2 test because 2 2(ln l) | x 2 with degrees of freedom equal to the number of restrictions. The degrees of freedom for our tests of homotheticity, homogeneity, and constant RTS are 3, 1, and 1, with critical values of 7.815, 3.841, and 3.841 at the 5 percent signicance levels. We calculate 2 2(ln l)55.491, 0.268, and 18.045 respectively. Therefore, the homothetic model is not rejected in favor of the general model, and the homogeneous model is not rejected in favor of the homothetic model. However, the constant RTS model is rejected in favor of the homogeneous model.22 These tests focus our attention on the homogeneous model. The model ts well: McElroys (1977) R 2 is 0.9985, and all but three model coefcients are signicant z at the 10 percent level or better. Further evidence of the performance of the homogeneous translog model is afforded by inspection of the tted share equations. For every observation, the tted shares fall in the interval (0, 1), as they should. In addition, the estimated cost function satises concavity with respect to advertising media prices for every observation.23 Finally, the tted own-price elasticities of demand, developed below, are negative for every observation. All this lends support for the homogeneous translog model. As a nal statistical test of the appropriateness of the homogeneous model we employ a Hausman (1978) specication test. Spencer and Berk (1981, p. 1079) point out that the Hausman test has a diffuse alternative hypothesis; it would reject the null hypothesis that the regression is correctly specied for a variety of reasons, including omission of relevant explanatory variables, errors in variables, inappropriate aggregation over time, simultaneity, incorrect functional form. In the Hausman test, we compare our coefcients for the homogeneous model estimated using NLITSUR, as reported in Table 2, against coefcients of the same

21 It is also possible for both ah and ag to be either signicant or insignicant; these possibilities are discussed in Davidson and MacKinnon (1981, 1993). 22 The hypothesis of constant RTS can also be tested directly from the homogeneous model. In that model, we estimate fq 5 1.2589. The t statistic for the null hypothesis that fq 2 1 5 0 is 3.52; so we reject the null hypothesis at the 1 percent signicance level. 23 We checked concavity of the cost function with respect to advertising prices by calculating the three principal minors of the Hessian matrix characterized by the element 2 Ct / Pj,t Pk,t for each year. Concavity of the cost function requires that these principal minors alternate from non-positive to non-negative to non-positive.

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1169

model estimated using nonlinear ordinary least squares (NLOLS). Under the null hypothesis of no misspecication, NLITSUR is consistent, asymptotically normal, and asymptotically efcient; while NLOLS will not be asymptotically efcient (Zellner, 1962, 1963; Zellner and Huang, 1962). Let bI and bO be the estimated column vectors from the NLITSUR and NLOLS regression systems, and let VI and VO be their covariance matrices. Then Hausmans m statistic is m 5 ( bI 2 bO )9(VI 2 VO )21 ( bI 2 bO ) where m | x 2 with 16 degrees of freedom, the number of coefcients estimated directly in the system of equations. The null hypothesis that the model is correctly specied is not rejected if m is statistically insignicant. We calculate m 5 15.6306 which is insignicant at the 5 percent level (the critical value is 26.296) and even at the 10 percent level (with a critical value of 23.542). This supports the homogeneous model. Because of the t and performance of the homogeneous translog model and because the model is supported by the Hausman test, calculations that follow will use coefcient estimates from this model.

7. Derivation of elasticities We are interested in substitutability among the advertising media. Uzawa (1962) showed that the Allen partial elasticities of substitution can be expressed as (see also Berndt and Wood, 1975)

s A 5 ( bii 1 S i2 2 Si ) /S i2 ; i 5 p, v, r ii
and

(7)

s A 5 ( bij 1 Si Sj ) /Si Sj ; i, j 5 p, v, r; i j ij

(8)

where (7) is the own-Allen and (8) is the cross-Allen elasticity. The Allen elasticities are symmetric, so s A 5 s A . ij ji An alternative asymmetrical measure of substitution originally developed by Morishima is related to the Allen elasticities (Blackorby and Russell, 1989).24 Let s M denote the Morishima elasticity. Then (Ball and Chambers, 1982) ij

24 Blackorby and Russell (1989) discussed why one should expect asymmetry in an elasticity of substitution. The elasticity of substitution measures the curvature along an isoquant when the ratio Pi /Pj varies due to a change in Pi . The direction of change is different when the ratio Pi /Pj varies due to a change in Pj , so we would expect the curvature to be different. Blackorby and Russell pointed out that the Morishima elasticities will be symmetric only in the case of CES production functions.

1170

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

s M 5 Sj (s A 2 s A ); i, j 5 p, v, r. ij ij jj

(9)

If i 5 j then the Morishima elasticity is zero. Only off-diagonal elements in a Morishima matrix may be non-zero. The Allen elasticities also can be used to calculate the own- and cross-price elasticities (eii and eij ) of the derived demand for inputs (Berndt and Wood, 1975) because

eij 5 Sj s A ; ij

i 5 p, v, r.

(10)

Like the Morishima elasticities, the price elasticities are asymmetric because Si Sj in general. Chambers (1988, p. 95) argued that the Allen elasticity yields no information beyond that given by the cross-price elasticity of demand because the Allen elasticity is merely the cross-price elasticity divided by the factor share of cost. Blackorby and Russell (1989) argued even more strongly that the Allen elasticity is uninformative. In contrast, they demonstrated that the Morishima elasticity is a meaningful and informative measure of substitutability. We agree with Blackorby and Russell. Although we use Allen elasticities to calculate Morishima and price elasticities, we do not report Allen elasticities in this paper. We do, however, offer them to the reader upon request. We measure scale economies along the expansion path as suggested by Hanoch (1975). To calculate the scale elasticity, hS , we rst calculate the cost elasticity, hC . The cost elasticity for the homogeneous model is simply 25

hC 5 (ln C) / (ln Q) 5 fq .
We dene the scale elasticity as the inverse of (11), so

(11)

hS 5 (ln Q) / (ln C) 5 (hC )21

(12)

With this denition, hS . 1 implies economies of scale in advertising while hS , 1 implies diseconomies of scale in advertising.26

25 For the general model, the cost elasticity would be hC 5 (ln C) / (ln Q) 5 fq 1 fqq (ln Q) 1 o i gqi (ln Pi ) for i 5 p, v, r. In the homogeneous model, the coefcients associated with the variables ln Q and ln Pi are zero, so hC 5 fq 26 An alternative denition of the scale elasticity is j S 5 1 2 hC (Christensen and Greene, 1976). With this denition, j S . 0 implies scale economies and j S , 0 implies scale diseconomies. Neither denition offers advantages over the other; j S . 0 is equivalent to hS . 1 and j S , 0 is equivalent to hS , 1.

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1171

8. Estimated substitution elasticities We calculate Morishima and price elasticities at two points. First, we calculate them at the means, as is common in the literature, because the translog cost function can be interpreted as a Taylor series expansion around the means of the variables and, as such, estimation at the means best approximates the true, but unknown, cost function. We also calculate the elasticities for our nal observation, the fourth quarter of 1993, because this will better represent the present and, therefore, will be useful in considering the policy implications of actions taken in the near future. In fact, the two points yield similar results. All elasticities are estimated using the coefcients of the homogeneous model of Table 2 and tted shares, which are functions of the coefcients and dependent variables, as suggested by Berndt (1991). Because the expected value of the lagged error term is zero, the tted shares for the homogeneous model, where gqi 5 0, are Si (P; ai , bi ) 5 ai 1

O b (ln P );
ij j j

i, j 5 p, v, r,

where Pj are either mean prices or prices for 1993:Q4, P is the price vector, bi is the vector of b coefcients, and denotes estimates from the homogeneous model. From Eq. (9), the estimated Morishima elasticities of substitution can be written as ij s M 5 bij /Si (P; ai , bi ) 2 b jj /Sj (P; aj , b j ) 1 1 with variance ij ij (s M / a s M / b )

O (s

M ij

ij / a s M / b )T

where a 5 ( ai aj ), b 5 ( bi b j ), o is the relevant portion of the variance covariance matrix, and T is the transpose operator.27 Taking the square root of the variance, we then form the t ratio. Price elasticities of demand and their associated standard deviations are calculated in a similar fashion. Morishima elasticities are presented in Table 3. The Morishima elasticity s M ij measures the curvature of the isoquant when adjustments are made in inputs i and j in response to a change in the price ratio Pi /Pj due to an increase in the price Pi

Many translog studies incorrectly treat shares as xed when estimating the variances of elasticities; they do not consider the effect of the parameters upon the shares. This results in biased estimates of the variances. In contrast, we estimate the variance as a Taylor series expansion. For a discussion of these issues, see Anderson and Thursby (1986).

27

1172

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

Table 3 Morishima elasticities of substitution a j5p j5v j5r 9.5267 d (3.6041) 7.9992 d (2.6664)

(A) Morishima elasticities at the mean values of independent variables i5p 0.1612 (0.8620) i5v 1.6887 b (1.2166) i5r 5.4099 c 4.2780 c (3.1260) (2.1758) (B) Morishima elasticities for the average rm, fourth quarter, 1993 i5p 0.1049 (0.9579) i5v 1.7380 b (1.2644) i5r 4.8528 b 3.7394 c (2.9598) (1.7319)

8.4873 d (3.0624) 6.8543 d (1.8779)

a Standard errors are in parentheses. Entries in the table are Morishima elasticities, s M , where rows ij denote the ith medium and columns denote the jth medium for i, j 5 p, v, r (print, television, radio). The own Morishima elasticity of substitution is not calculated for i 5 j. b Signicant at the 10 percent level, one-tailed test. c Signicant at the 5 percent level, one-tailed test. d Signicant at the 1 percent level, one-tailed test.

(Blackorby and Russell, 1989, p. 885). As discussed previously, this will typically be different from the curvature moving in the other direction, when Pi /Pj changes due to an increase in the price Pj . For instance, consider the substitution between television and print media using the mean values of the variables. If the price of TV advertising increases, the estimate for s M in Table 3A (1.69, signicant at the vp 10 percent level) indicates that rms are able to substitute print for TV relatively easily.28 On the other hand, as the price of print media advertising increases, the results in Table 3A suggest that substitution into TV is not so readily accom pv plished ( s M 5 0.16 in Table 3A, and insignicantly greater than 0). This asymmetry may result from the fact that television accounts for 81 to 85 percent of tted advertising expenses while print accounts for only 9 to 13 percent of these expenses. Therefore, because television is already heavily saturated relative to print, increases in the price of print advertising induce beer producers to substitute
28

As a basis of comparison, the simple elasticity of substitution for the two-input Cobb-Douglas production function is unity. Because the Cobb-Douglas function has a long history in economic analysis, it seems reasonable to use unity as a benchmark. Blackorby and Russell (1989) showed that, unlike the Allen elasticity, the Morishima elasticity is unity for two inputs that enter the production function in a Cobb-Douglas form. As the Morishima elasticity approaches innity, the two inputs become perfect substitutes.

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1173

pr more heavily into radio ( s M 5 9.53 in Table 3A, signicant at the 1 percent level). The results in Table 3B suggest similar conclusions for the fourth quarter of 1993. In general, Table 3A and B suggest that the media are fairly good substitutes. While our elasticity estimates are based on marginal changes in advertising, they may hold more generally. To the extent that this is true, it suggests two policy implications. First, partial bans may be ineffective because rms banned from advertising in some media may switch to other media as outlets for their advertising. This may be particularly relevant for broadcast bans because print advertising appears to be a fairly good substitute for both TV and radio advertising: in Table 3B, where substitution elasticities are calculated for our latest vp rp data, s M and s M are 1.74 and 4.85 and are signicant at the 10 percent and 5 percent signicance levels, respectively. Second, with respect to mergers in the television and radio media, antitrust agencies perhaps need not be too concerned that the owners of these media outlets will be able to signicantly increase the price of advertising because advertisers could switch to print advertising. Similar vr ly, if the price of TV advertising increased, rms can switch into radio ( s M 5 6.85, signicant at the 1 percent level); and if the price of radio advertising rv increased, rms can switch into TV ( s M 5 3.74, signicant at the 5 percent level). pr The highest substitution elasticity is from print into radio ( s M 5 8.49, signicant at the 1 percent level).
Table 4 Price elasticities of demand a j5p j5v j5r 1.7811 b (1.2453) 0.2537 b (0.1590) 27.7456 d (2.6192)

(A) Price elasticities at the mean values of the independent variables i5p 21.6994 b 20.0817 (1.1607) (1.1569) i5v 20.0107 20.2430 b (0.1593) (0.1595) i5r 3.7105 b 4.0350 b (2.6165) (2.6004) (B) Price elasticities for the average rm, fourth quarter, 1993 i5p 21.7557 b 20.1428 (1.2083) (1.2065) i5v 20.0177 20.2477 b (0.1584) (0.1692) i5r 3.0971 c 3.4918 c (1.8582) (1.8431)
a

1.8984 b (1.2129) 0.2654 c (0.1555) 26.5889 d (1.8591)

Standard errors are in parentheses. Entries in the table are price elasticities of demand, eij , where rows denote the ith medium and columns denote the jth medium for i, j 5 p, v, r (print, television, radio). Own-price elasticities are on the main diagonal. b Signicant at the 10 percent level, one-tailed test. c Signicant at the 5 percent level, one-tailed test. d Signicant at the 1 percent level, one-tailed test.

1174

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

Estimates of price elasticities of demand ( eij ) for the mean of the observations and for the fourth quarter of 1993 are presented in Table 4. Estimated own-price elasticities of demand ( eii ) are shown along the main diagonals of Table 4A and B. As required by the law of demand, the own-price elasticities are all negative; and all are signicant at the 10 percent level or better. At the relevant points on their demand functions, demand for television advertising is relatively inelastic, while demand for radio and print media are relatively elastic. The inelasticity of TV advertising demand implies that market power is not being exerted in the setting of TV ad rates, because rms with market power have incentive to price in the elastic portion of their demand curve. The relative elasticity of demand for print and radio may be due, in part, to the fact that they account for smaller portions of rms advertising budget. However, they also may be due, in part, to market power enjoyed by the rms that sell print and radio advertising space. Nevertheless, with the exception of the cross-price elasticities between TV and print, these elasticities support our ndings discussed above that the media are substitutes, so that market power is probably not extreme among the suppliers of advertising space. Both cross-price elasticities involving TV and print are negative (in Table 4A, evp 5 2 0.01 and epv 5 2 0.08), but both are highly insignicant. If the true values were one standard deviation above these estimates, both would be positive (0.15 and 1.08); we cannot place much condence in these estimates.

9. Economies of scale in advertising From Eq. (11), the cost elasticity, hC , is estimated by fq in the homogeneous model. It is the same for all observations and for the mean of the variables because this elasticity is constant. From Table 2, hC is 1.2589 with a standard error of 0.0735, signicant at the 1 percent level. The scale elasticity is hS 5 ( hC )21 5 0.7943 with standard error 0.0464, signicant at the 1 percent level. This suggests diseconomies of scale in beer advertising, at least over the advertising levels of rms in our sample (see Table 1). This contradicts some earlier results such as Peles (1971) and Brown (1978); see also Krouse (1990, p. 500), Martin (1994, pp. 245247), and Berndt (1991, p. 415). However, Thomas (1989) also found scale diseconomies in advertising for cigarettes and soft drinks. In addition, in an earlier study, Simon (1965) reported a preponderance of evidence for a number of different goods that the rst ad is the most efcient, and additional repetitions do less and less work. 29 Our nding of diseconomies of scale in beer advertising implies that large-scale advertising, such as the levels employed by rms in our sample, is not useful in building barriers to entry into the beer market.

Some studies (e.g., Tremblay, 1987) confound economies of scale in production with economies of scale in advertising, so we cannot compare our results to these studies.

29

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1175

Diseconomies of scale are certainly plausible. It is reasonable, and eminently rational, that advertising efforts should rst be made through media outlets that would reach the largest audiences (e.g., national magazines and radio and TV shows aired nationwide). Then, as advertising efforts increase, media with less reach would be used (such as regional magazines and spot radio and TV markets). It follows that marginal audiences would diminish in size as advertising efforts increase. Finally, we compare our results to Bresnahans (1984). In the section of his study that used data for six industries, Bresnahan deemed his results concerning scale economies inconclusive because of the insignicance of proxy variables he used in place of quantity (time and four-rm concentration ratios). He stated (p. 147) that his evidence could support either increasing returns to scale or a lessening of diseconomies of scale over time. In the part of his study that considered beer, Bresnahan found some support for economies of scale. However, our results are not directly comparable to his because his methodo1ogy is different from ours. He regressed the advertising / sales ratio on a number of independent variables, including a price index for all advertising messages. He then arrived at his conclusion based upon the sign associated with the price index coefcient in three alternative specications of the advertising / sales ratio. In contrast to Bresnahan, who acknowledged that his regressions to measure scale economies were somewhat ad hoc,30 we follow Hanoch (1975) by measuring the scale elasticity along the expansion path using parameters from the cost function.

10. Conclusion This study focused upon the substitutability of advertising media and advertising economies of scale using a translog model and rm-level data for beer manufacturers over the period 19831993. We estimated Morishima elasticities of substitution and price elasticities and found that all advertising media are substitutes in promoting sales. We found strong substitution possibilities from TV into both print and radio, from radio into both print and TV, and from print into radio. In addition, we found evidence of diseconomies of scale in advertising. Aside from academic interest, these two ndings suggest three policy implications. First, recent events in the U.S. have refocused attention upon the possibility of a partial ban on alcohol advertising, both for liquor and for beer. A ban on TV and

Bresnahan (1984) regressed the advertising / sales ratio on quantitative measures of the rms environment, but not on any choice variables of the rm (p. 154). He further stated, [i]t is not obvious what the determinants of [advertising / sales] should be and presented three versions of the regression. In contrast, our approach allows the cost function to determine the scale elasticity.

30

1176

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

radio advertising could be effective in reducing alcohol-related problems if print media are not good substitutes for TV and radio. However, our results suggest that beer producers can substitute fairly easily into print advertising, so a partial ban may be ineffective if the effects are noticeable at all. To the extent that the implications of our results, which hold for marginal changes, are indicative of larger changes, beer producers probably would respond to a ban on TV and radio advertising by increasing advertising in printed media. The fact that U.S. liquor producers voluntarily avoided broadcast advertising from 1936 until the recent advertising campaign by Seagram supports this possibility: liquor producers evidently have viewed other advertising media as close substitutes for broadcast advertising. Moreover, our results support evidence from research on cigarette advertising that suggests that partial bans have not been highly effective in reducing cigarette consumption (Seldon and Doroodian, 1989).31 This implication may hold for goods other than alcohol and cigarettes. For instance, consider childrens toys. Restrictions on TV advertising directed at children may be ineffective because advertising appears in childrens publications. While we do not examine other markets, it is clear that the effectiveness of a partial ban in any market will be mitigated by media substitutability.32 Second, regulatory agencies in the U.S. and Mexico have been concerned that increased concentration in TV and radio markets may result in the ability of large TV and radio corporations to raise advertising rates. Our estimated own-price elasticities suggest that TV corporations currently lack market power in setting advertising rates, while print and radio media owners may have some market power. However, our estimated substitution elasticities suggest that such market power is, and is likely to remain, limited. More succinctly, the results of this study imply that mergers in the entertainment industry will not lead to signicant market power in setting advertising rates. Third, economies of scale in advertising have implications for the ability of advertising to raise barriers to entry into the beer market. In contrast to concerns that economies of scale in advertising can insulate incumbent rms from the threat of entry, we nd diseconomies of scale, at least in the range of advertising of rms in our sample. This suggests that large-scale advertising may not create a barrier to

An exception is a recent paper by Tremblay and Tremblay (1997). They present weak evidence that television and radio are the most effective media for promoting [cigarette] sales (pp. 1011). They also suggest that the broadcast advertising ban indirectly reduced consumption. Their statistical results suggest that the broadcast advertising ban increased market power among cigarette companies leading them to increase prices. In response to higher prices, consumers reduced consumption. 32 Media substitution may change the types of consumers who purchase a good. While this study shows that the rm may maintain sales with media substitution, the composition of consumers could be different. This issue is beyond the scope of this study.

31

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1177

entry due to scale economies, at least into the beer market. This implication may hold for other markets as well.33 We offer one nal word, with reference to substitution among, and relative effectiveness of, the various advertising media. This study suggests that sales can be maintained using different combinations of advertising media. Therefore, contrary to the claim of Marshall McLuhan (1964), it appears that the message is more important than the medium.

Acknowledgements The authors thank Robert Coen of McCann-Erickson for providing media cost data, Jane Darling of The University of Texas at Dallas Library for assistance in data collection, and Devon Herrick for research assistance. We thank Joseph E. Harrington, Jr., Wim P.M. Vijverberg, and an anonymous referee for constructive comments and suggestions. The usual disclaimer applies. Authorship is shared equally. An earlier version of this paper was presented at the 25th Annual Conference of the European Association for Research in Industrial Economics in Copenhagen, August 1998.

References
Albion, M.S., Farris, P.W., 1981. The Advertising Controversy: Evidence On the Economic Effects of Advertising. Auburn House Publishing, Boston. Anderson, R.G., Thursby, J.G., 1986. Condence intervals for elasticity estimators in translog models. Review of Economics and Statistics 68, 647656. Ashley, R., Granger, C.W.J., Schmalensee, R., 1980. Advertising and aggregate consumption: An analysis of causality. Econometrica 48, 11491167. Associated Press, 1996. Clinton criticizes distillers for ending liquor-ad ban. The Dallas Morning News, 10 November 1996, 18A. Ball, V.E., Chambers, R.G., 1982. An economic analysis of technology in the meat products industry. American Journal of Agricultural Economics 64, 699709. Bass, F.M., Leone, R.P., 1983. Temporal aggregation, the data interval bias, and empirical estimation of bi-monthly relations from annual data. Management Science 29, 111. Beatty, S.G., 1996a. Seagram outs ban on TV ads pitching liquor. The Wall Street Journal, 11 June 1996, B1 and B6. Beatty, S.G., 1996b. Liquor industry is divided over use of TV ads. The Wall Street Journal, 12 June 1996, B1 and B3.

33 Sutton (1991) points out that the necessity of advertising may pose a barrier to entry because advertising represents a cost that must be sunk in order to compete in a market. Because of this, advertising may erect a barrier to entry even if there are no economies of scale in advertising.

1178

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

Beatty, S.G., 1996c. TV beer ads may be caught in a backlash. The Wall Street Journal, 14 June 1996, B1 and B5. Beatty, S.G., 1996d. New rules are sought on ad marketing. The Wall Street Journal, 25 June 1996, B11. Beatty, S.G., 1996e. Liquor industry votes to rescind ban on TV ads. The Wall Street Journal, 8 November 1996, B1. Beatty, S.G., 1998. FTC eyes liquor ads kid-appeal. The Wall Street Journal, 7 August 1998, B1 and B5. Brewers Almanac, 1994, Beer Institute, Washington, DC. Berndt, E.R., 1991. The Practice of Econometrics: Classic and Contemporary. Addison-Wesley Publishing, Reading, MA. Berndt, E.R., Savin, N.E., 1975. Estimation and hypothesis testing in singular equation systems with autoregressive disturbances. Econometrica 43, 937957. Berndt, E.R., Wood, D.O., 1975. Technology, prices, and the derived demand for energy. The Review of Economics and Statistics 57, 259268. Beverage Producers Association, Various Issues. Beverage Industry. Stagnito Publishing, Northbrook, IL. Blackorby, C., Russell, R.R., 1989. Will the real elasticity of substitution please stand up? (A comparison of the Allen / Uzawa and Morishima elasticities). The American Economic Review 79, 882888. Boyd, R.G., Seldon, B.J., 1990. The eeting effect of advertising. Economics Letters 34, 375379. Bresnahan, T.F., 1984. The demand for advertising by medium: Implications for the economies of scale in advertising. In: Ippolito, P.M., Scheffman, D.T. (Eds.). Empirical Approaches to Consumer Protection Economics. Federal Trade Commission, Washington, DC, pp. 135163. Brown, R.S., 1978. Estimating advantages to large-scale advertising. The Review of Economics and Statistics 60, 428437. Caves, D.W., Christensen, L.R., Swanson, J.A., 1981. Productivity growth, scale economies, and capacity utilization in U.S. railroads, 19551974. The American Economic Review 71, 9941002. Chambers, R.G., 1988. Applied Production Analysis: A Dual Approach. Cambridge University Press, Cambridge, UK. Christensen, L.R., Greene, W.H., 1976. Economies of scale in U.S. electric power generation. Journal of Political Economy 84 (Part 1), 655676. Competitive Media Reporting, Annual Issues 198393. Ad $ Summary. Competitive Media Reporting, New York. Davidson, R., MacKinnon, J.G., 1981. Several tests for model specication in the presence of alternative hypotheses. Econometrica 49, 781793. Davidson, R., MacKinnon, J.G., 1993. Estimation and Inference in Econometrics. Oxford University Press, New York. Eckard, Jr. E.W., 1991. Competition and the cigarette advertising ban. Economic Inquiry 29, 119133. Ekelund, R.B., Saurman, D.S., 1988. Advertising and the Market Process: A Modern Economic View. Pacic Research Institute for Public Policy, San Francisco. Elzinga, K.G., 1995. Beer. In: Adams, W., Brock, J. (Eds.), The Structure of American Industry, 9th edn. Prentice-Hall, Englewood Cliffs, NJ, pp. 119151. Fritschler, A.L., 1989. Smoking and Politics, 4th edn. Prentice Hall, Englewood Cliffs, NJ. Greene, W.H., 1993. Econometric Analysis. Macmillan Publishing, New York. Grossman, M., Sindelar, J.L., Mullahy, J., Anderson, R., 1993. Policy watch: Alcohol and cigarette taxes. Journal of Economic Perspectives 7, 211222. Hanoch, G., 1975. The elasticity of scale and the shape of average costs. The American Economic Review 65, 492497. Hausman, J.A., 1978. Specication tests in econometrics. Econometrica 46, 12511271. Hay, D.A., Morris, D.J., 1991. Industrial Economics and Organization: Theory and Evidence. Oxford University Press, New York.

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

1179

Ingersoll, B., 1996. FTC opens investigation of TV alcohol advertising. The Wall Street Journal, 27 November 1996, A3 and A6. Jewell, R.T., Brown, R.W., 1995. Alcohol availability and alcohol-related motor vehicle accidents. Applied Economics 27, 759765. Jung, C., Seldon, B.J., 1995. The degree of competition in the advertising industry. Review of Industrial Organization 10, 4152. Krouse, C.G., 1990. Theory of Industrial Economics. Basil Blackwell, Cambridge, MA. Lee, B., Tremblay, V.J., 1992. Advertising and the U.S. market demand for beer. Applied Economics 24, 6976. Leone, R.P., 1995. Generalizing what is known about temporal aggregation and advertising carryover. Marketing Science 14, G141G150. Martin, S., 1994. Industrial Economics: Economic Analysis and Public Policy, 2nd edn. Macmillan, New York. McAuliffe, R.E., 1987. Advertising, Competition, and Public Policy: Theories and New Evidence. Lexington Books, Lexington, MA. McCullough, B.D., Waldron, T., 1998. The substitutability of network and national spot television advertising. Quarterly Journal of Business and Economics 37, 315. McElroy, M.B., 1977. Goodness of t for seemingly unrelated regressions. Journal of Econometrics 6, 381387. McLuhan, M., 1964. Understanding Media. McGraw-Hill, New York. Peles, Y., 1971. Economies of scale in advertising beer and cigarettes. Journal of Business 44, 3237. Pope, K., 1997. FCC chairman urges broadcasters to form voluntary policy on liquor ads. The Wall Street Journal, 9 April 1997, B6. Saffer, H., 1991. Alcohol advertising bans and alcohol abuse: An international perspective. Journal of Health Economics 10, 6579. Saffer, H., 1993. Advertising under the inuence. In: Hilton, M.E., Bloss, G. (Eds.), Economics and the Prevention of Alcohol-Related Problems. National Institute on Alcohol Abuse and Alcoholism, Rockville, MD, pp. 125140. SAS / ETS Users Guide, 1988, Version 6. SAS Institute, Cary, NC. Scherer, F.M., Ross, D., 1990. Industrial Market Structure and Economic Performance, 3rd edn. Houghton Mifin, Boston. Seldon, B.J., Bullard, S.H., 1992. Input substitution, economies of scale, and productivity growth in the U.S. upholstered furniture industry. Applied Economics 24, 10171025. Seldon, B.J., Doroodian, K., 1989. A simultaneous model of cigarette advertising: Effects on demand and industry response to public policy. The Review of Economics and Statistics 71, 673677. Seldon, B.J., Jung, C., 1993. Derived demand for advertising messages and substitutability among the media. The Quarterly Review of Economics and Finance 33, 7186. Simon, J.L., 1965. Are there economies of scale in advertising? Journal of Advertising Research 5, 1520. Smart, R.G., 1988. Does alcohol advertising affect overall consumption? A review of empirical studies. Journal of Studies on Alcohol 49, 314323. Spencer, D.E., Berk, K.N., 1981. A limited information specication test. Econometrica 49, 1079 1085. Sutton, J., 1991. Sunk Costs and Market Structure: Price Competition, Advertising, and the Evolution of Concentration. The MIT Press, Cambridge, MA. Thomas, L.G., 1989. Advertising in consumer good industries: Durability, economies of scale, and heterogeneity. Journal of Law and Economics 32, 164194. Tremblay, C.H., Tremblay, V.J., 1995. Advertising, price, and welfare: Evidence from the U.S. brewing industry. Southern Economic Journal 62, 367381. Tremblay, C.H., Tremblay, V.J., 1997. The effect of advertising restrictions on competition and cigarette consumption in the United States. Oregon State University, mimeo dated July 1997.

1180

B. J. Seldon et al. / Int. J. Ind. Organ. 18 (2000) 1153 1180

Tremblay, V.J., 1987. Scale economies, technical change, and rm-cost asymmetries in the U.S. brewing industry. The Quarterly Review of Economics and Business 27, 7186. Uzawa, H., 1962. Production functions with constant elasticities of substitution. Review of Economic Studies 29, 291299. Wagenaar, A.C., 1993. Minimum drinking age and alcohol availability to youth: Issues and research needs. In: Hilton, M.E., Boss, G. (Eds.), Economics and the Prevention of Alcohol-Related Problems. National Institute on Alcohol Abuse and Alcoholism, Rockville, MD, pp. 175200. Waterson, M., 1984. Economic Theory of the Industry. Cambridge University Press, Cambridge, UK. Young, D.J., 1993. Alcohol advertising bans and alcohol abuse: comment. Journal of Health Economics 12, 213228. Zellner, A., 1962. An efcient method for estimating seemingly unrelated regressions and tests for aggregation bias. Journal of the American Statistical Association 57, 348368. Zellner, A., 1963. Estimators for seemingly unrelated regression equations: Some exact nite sample estimates. Journal of the American Statistical Association 58, 977992. Zellner, A., Huang, D.S., 1962. Further properties of efcient estimators for seemingly unrelated regression equations. International Economic Review 3, 300313.

Das könnte Ihnen auch gefallen