Was Mr. Hewlett Right? Mergers, Advertising and the PC Industry. Michelle Sovinsky Goeree 1

Transcription

1 Was Mr. Hewlett Right? Mergers, Advertising and the PC Industry Michelle Sovinsky Goeree 1 February, 2009 Abstract In markets characterized by rapid change consumers may not know every available product. Advertising allows rms to inform consumers about their products, but rms pro tmaximizing advertising choices are not necessarily welfare maximizing. Ignoring the consequences of limited information and rms strategic incentives for providing information in merger analysis could lead antitrust authorities to approve a merger that has negative consequences for welfare. I use the parameter estimates from a limited information model in Goeree (2008) to examine the role of strategic information provision via advertising and examine the implications for antitrust merger policy. To do so, I simulate post-merger price and advertising equilibria. I decompose the post-merger change in prices into changes due to increased concentration and changes due to strategic provision of information. The results indicate advertising can be used to increase market power when consumers have limited information, which suggests revisions to the current model used to access the impact of mergers in antitrust cases. JEL Classi cation: L15, D12, M37, L44, D83 Keywords: merger analysis, informative advertising, discrete-choice models, product di erentiation, structural estimation 1 University of Southern California, Economics Department (goeree@usc.edu). This research has bene ted from comments from seminar participants at Claremont McKenna, the Federal Trade Commission, EARIE meetings (Porto), IIOC meetings (Boston), and discussions with Ulrich Doraszelski and Michael Waterson. I am grateful to Gartner Inc. for making the data available and for nancial support from the University of Virginia s Bankard Fund for Political Economy.

2 1 Introduction On May 7, 2002 Hewlett-Packard (HP) Company launched the new HP with an ad titled We are Ready. The new HP is a result of a merger with Compaq Computer Corporation, the largest ever in the information technology sector. The $19 billion deal drew a lot of media attention for a number of reasons. Investors and rival rms were interested in its impact on shares and pro ts. Consumers were interested in the e ect on prices. Regulators were interested in its implications for competition in an already concentrated industry. Originally proposed in June 2001, the merger prompted a bitter battle between Hewlett and Packard family interests and corporate executives. It was ultimately approved by a slim majority of shareholders (only 3%). Many HP shareholders opposed the deal because they thought the time lost in absorbing Compaq and incorporating cost synergies would distract from winning new orders at a time when the market was slowing. Walter Hewlett, whose father cofounded HP, initiated a court battle against HP arguing the merger would result in lost pro ts in the long run. As further evidence of his conjecture, Hewlett pointed to his competitors, We believe that HP stockholders should be concerned when competitors, like SUN, Dell, and IBM don t object to a transaction that is supposed to add value to HP. Meanwhile, the Federal Trade Commission (FTC) voted unanimously to approve the merger. Likewise the European Commission approved it without placing any conditions on the two companies saying, A careful analysis of the merger... has shown that HP would not be in a position to increase prices and that consumers would continue to bene t from su cient choice and innovation. 2 Currently, antitrust authorities evaluate the impact of mergers under the assumption that consumers know all products for sale when making their purchases. This assumption is questionable, especially when applied to markets characterized by rapid change, such as the personal computer (PC) industry. Advertising allows rms to inform consumers about their products, but rms pro t-maximizing advertising choices are not necessarily welfare maximizing and can result in increased market power at the expense of consumers. Indeed Goeree (2002, 2008) shows that assuming full-information in a market characterized by rapid change generates estimates of demand curves that are biased towards being too elastic. This result is of particular importance when considering the welfare impact of mergers, the primary interest of US antitrust authorities. Ignoring the consequences of limited information and rms strategic incentives for advertising in merger analysis could lead antitrust authorities to approve a merger that has negative consequences for welfare. 2 HP-Compaq Merger Wins European Approval, NewsFactor Network, Feb.1,

3 The goal of this empirical work is twofold: (1) to estimate the e ect on pro ts and consumer welfare from mergers when consumers have limited information and (2) to examine the role of strategic information provision via advertising and the resulting implications for antitrust policy. The methodology used to evaluate the impact of mergers is based on previous work, 3 but allows for limited information and strategic choices of advertising. I use the estimated parameters from a structural model of limited information presented in Goeree (2008) to simulate post-merger equilibrium price and advertising levels. I calculate the e ect of mergers on the pro ts of merging and non-merging rms as well as the cost-synergies necessary to o set losses. I decompose the post-merger change in prices and markups into the changes due to increased concentration and the changes due to the in uence of information provision. The post-merger equilibrium results indicate advertising can be used to increase market power, which suggests revisions to the current model used by antitrust authorities to determine market power in antitrust cases. I examine both the HP-Compaq merger and a hypothetical merger between IBM and Dell over a period of intense growth in the PC market, Considering a merger between IBM and Dell is of interest for two reasons. First, IBM is the world leader in sales of non-pc s, while Dell s focus is more on the sale of PC s. As a result, cost synergies would likely be lower than those between HP and Compaq. Secondly, and more directly related to the topic of this research, IBM and Dell have very di erent ad-to-sales ratios. IBM is a high-advertising intensity rm (ad-to-sales ratio over 20%), while Dell is a low-intensity rm (ad-to-sales ratio under 3%). cost synergies and are closer in their ad-sales concentration. In contrast, HP and Compaq are thought to have more Comparison of the outcomes from the two very di erent mergers yields insight into the role that advertising plays in this industry and how rms use it when industry concentration increases. Related literature forthcoming... The remainder of the paper is organized as follows. data used in estimation. In the next section, I discuss the In section 3, I describe the model used in the counterfactual merger simulations. In section 4, I discuss the estimation technique and present parameter estimates. In section 5, I discuss the impact of mergers on welfare and pro ts, the role that advertising plays as concentration increases, and the implications for antitrust policy. The nal section concludes. 3 For example, see Baker and Bresnahan, 1985; Berry and Pakes, 1993, Hausman, Leonard and Zona, 1994; and Nevo, While the merger is supposed, IBM and Dell entered into a $16 billion cross-licensing agreement in The agreement called for broad patent cross-licensing between the two rms and collaboration in the development of product technology. 2

4 2 Data The data come from three primary sources. Product-level data are from Gartner Inc. and consist of prices, market shares, and product characteristics of all PCs sold to home users between 1996 and Home market sales accounts for over 30% of all PCs sold. I have data on ve main PC attributes: rm (e.g. Dell), brand (e.g. Latitude LX), form factor (e.g. desktop), CPU type (e.g. Pentium II), and CPU speed (MHz). I de ne a model as a rm, brand, CPU type, CPU speed, form factor combination. 5 Treating a model/quarter as an observation, the total sample size is Percentage Dollar Average Annual Manufacturer Home Market Share Ad Ad to Sales Median Price Expend Ratio Home Sector Industry 3.4% $2,239 Top 6 Firm $ % $2,172 Acer $ % $1,708 Apple $ % $1,859 AST Compaq $ % $2,070 Dell $ % $2,297 Gateway $ % $2,767 Hewlett Packard $ % $2,203 IBM $1, % $2,565 Micron NEC 3.22 Packard Bell Packard Bell NEC $ % $2,075 Texas Instruments 1.40 Notes: In 1997 three mergers occurred :Packard Bell, NEC,ZDS; Acer,Texas Instr.; Gateway, Advanced Logic Research. Ad expenditures (in $M) and ad to sales ratios are annual averages and are from LNA and include all sectors (home, business, education, government). Table 1: Summary Statistics Table 1 presents descriptive statistics for the home market sector. The top six rms account for over 69% (71%) of the dollar (unit) home market share on average. The major market players did not change over the period, although there was signi cant change in some of their market shares. In estimation I consider the top ten rms and ve others (AST, 5 The data I use consist of attributes which cannot be easily altered or enhanced after purchase. The data allow for a very narrow model de nition. For example, the Compaq Armada 6500 and the Armada 7400 are two separate models. Both have Pentium II 300/366 processors, 64 MB standard memory, 56KB/s modem, an expansion bay for peripherals, and full-size displays and keyboards. The 7400 is lighter, although somewhat thicker, and it has a larger standard hard drive, and more cache memory. In both models the hard drive and memory are expandable up to the same limit. In addition, the Apple Power Macintosh Power PC /200 desktop and deskside are two separate models. They di er only in their form factor. 3

5 AT&T / NCR, DEC, Epson, and Texas Instruments) to make full use of micro-purchase data discussed below. The included rms are Acer, Apple, AST, AT&T, Compaq, Dell, Epson, Gateway, HP, IBM, Micron, NEC, Packard Bell, and Texas Instruments. Together these 15 rms account for over 85% (83%) of the dollar (unit) home market share on average. This study considers the period when the PC industry was experiencing tremendous growth. In 1996 Packard Bell was a 4.5 billion company and the largest PC manufacturer in the US. 6 Compaq passed Packard Bell in mid 1996 and price pressure from Compaq and emachines, along with poor showings in consumer satisfaction surveys, made it di cult for the company to remain pro table. In 1997, three mergers occurred: Packard Bell, NEC, and ZDS; Acer and Texas Instruments; and Gateway and Advanced Logic Research. After this period, there was a slowdown in the PC market. Demand in the home market sector (as well as other sectors) declined. In part because there was not as much of a need to upgrade as often since the PCs were so well made and, in part due to the slump in the economy. Immediately after 1998, Compaq merged with DEC, which proved to be a mistake for Compaq in that they never fully recovered their pre-merger market position. It is this Compaq with which HP merged in I combine the product level data with advertising data from Competitive Media Reporting s LNA/ Multi-Media. These are quarterly advertising expenditures across media by rm with some brand level information. The expenditures are reported for ten media from which I construct four categories: newspaper, magazine, television, and radio. 7 There are two points to make regarding the advertising data. First, they are not broken down by sector. Hence, some of the advertising I observe is for non-pcs intended for non-home consumers. Second, PC rms may advertising their products in groups (using a combination of product-speci c advertising and group advertising with groups of varying sizes). For example, in 1996 one of Compaq s advertising campaigns involved all Presario brand computers (of which there are 12). In Goeree (2008) I show how to deal with this issue by constructing a measure of ad expenditures by product that incorporates all advertising done for the product. I construct e ective product advertising by adding observed product-speci c expenditures to a weighted average of all group expenditures for that product where the 6 Packard Bell was an American radio manufacturer. The Packard Bell Company has no association with Hewlett Packard. 7 The Magazine medium includes Sunday magazines; TV includes network, spot, cable or syndicated TV; Radio includes network and spot radio. There are many zero observations for outdoor advertising, and so I include it in the radio medium. Internet advertising was not so prevalent over this time period. 4

6 weights are estimated. 8 As can be seen from Table 1, there is much variation in advertising expenditures across rms. For instance, in 1998, the majority of the industry expenditures are by IBM, resulting in an ad-to-sales ratio of over 20 percent. While the industry average ad-to-sales ratio is much lower, about 3%. Excluding IBM s expenditures, the remaining top rms spend an average of 6.5% of their revenue on advertising. In contrast, Compaq s ad-to-sales ratio is only 2.4%. Variable Description Mean Std. Dev. male white age (years) to50 (=1 if 30<age<50) education (years) married household size employed income ($) inclow (=1 if income<$60,000) inchigh (=1 if income>$100,000) own pc (=1 if own a PC) pcnew (=1 if PC bought in last 12 months) cable (=1 if receive cable) hours cable (per day) hours non cable (per day) hours radio (per day) magazine (=1 if read last quarter) number magazines (read last quarter) weekend newspaper (=1 if read last quarter) weekday newspaper (=1 if read last quarter) Notes: Number of observations in survey is 39,931. Sample size is 13,400. Media exposure summary statistics are based on reports published by Simmons. Table 2: Simmons Data Descriptive Statistics I also use data from an annual micro-level survey of about 20,000 households collected by Simmons Market Research. The Simmons data contain information on consumer characteristics and PC purchases (although only the rm from which the individual purchased). Simmons data also include information on the media habits of the consumers including how 8 Speci cally, let G j be the set of all possible product groups that include model j. Let ad H be total e ective advertising expenditures for H 2 G j. De ne ad H ad H jhj : Then e ective advertising expenditures for product j are given by ad j = X 1 ad H + 2 ad 2 H H2G j where the sum is over the di erent groups that include product j: The parameters 1 and 2 are estimated if the product is advertised in a group, otherwise 1 is restricted to one and 2 to zero. 5

7 often individuals watched TV, read magazines, etc. The information on media exposure is valuable in that it allows me to link consumer characteristics to media with variation across media categories. I use two years of the Simmons survey from (1998 were not publicly available). Table 2 presents descriptive statistics for the Simmons data. Finally, I use also data from the Consumer Population Survey to de ne the distribution of consumer characteristics because the Simmons data are not available over all the years. 9 Market shares are unit sales of each model divided by market size. More detail about the various datasets, their construction, and descriptive statistics can be found in Goeree (2008). Every year over 200 new PCs are made available from the top 15 rms alone (Gartner, 2002). Due to the large number of product introductions and the competitive nature of the industry, it is has become increasingly important for rms to advertise to inform consumers about their products. As Figures 1 and 2 illustrate, prices dropped from an average of close to $2600 in early 1997 to just above $1500 in 1999, while advertising expenditures grew by over $0.5 billion. Advertising has been an important dimension of competition in this industry since its beginnings. Between 1995 and 1999, advertising expenditures grew by nearly 100% to $2.3 billion. In 1998 over 36 million PCs were sold, generating over $62 billion in sales of which $2 billion was spent on advertising. $2, Average Price (98$) $2,500 $2,300 $2,100 $1,900 $1, Total Units (in millions) $1, Average Price Total Units Sold Figure 1 9 I drew a sample of 3,000 individuals from the March CPS for each year. Quarterly income data were constructed from annual data and were de ated using the Consumer Price Index from BLS. A few households reported an annual income below $5000. These households were dropped from the sample. Examination of the Simmons data indicate that purchases were made only by households with annual income greater than $5000, therefore eliminating very low income households should not a ect the group of interest. 6

8 Millions of 98 Dollars $2,500 $2,250 $2,000 $1,750 $1,500 $1, Figure 2: Advertising Expenditures Competition and technological improvement has continued to spur innovation, and the PC industry has seen numerous product introductions in the past years. Due to the frequency with which new products are brought into the market, consumers may not be aware of all products o ered. In the next section I present a model of limited consumer information, where advertising can be used by rms to inform consumers about their products. 3 Model When the FTC or Department of Justice (DOJ) analyzes the competitive impact of mergers they rst determine pre-merger prices and markups. The most obvious way to determine markups over marginal costs would be to gather cost and price data for the industry of interest and directly compute markups. However, cost data are di cult to come by and, if they are found, are often proprietary. The approach in this paper follows recent studies of di erentiated products, such as Berry, Levinsohn, and Pakes (1995, 2004) (hereafter BLP), Nevo (2000), and Petrin (2002). I estimate a demand system and use the estimated demand elasticities to compute markups. I use the Nash rst order conditions to back out marginal costs as price minus markup. Given marginal costs, demand, and a notion of equilibrium I compute the equilibrium prices that would result under various ownership structures brought about through mergers. Unlike previous merger studies, this study i) incorporates limited consumer information when examining the post-merger decisions rms make regarding prices and ii) allows for changes in advertising post-merger, where advertising may impact the choice set of consumers. 7

9 3.1 Model of Limited Consumer Information Individual i = 1; :::; N chooses from j = 1; :::; J products at time t = 1; :::; T. A product pertains to a speci c PC model de ned as a rm-brand- CPU type-cpu speed-form factor combination. The indirect utility consumer i obtains from product j at time t is given by u ijt = ln(y it p jt ) + x 0 j it + jt + ijt (1) The characteristics of product j are represented by (p jt ; x j ; jt ); these are price, non-price observed characteristics (such as CPU speed, laptop and Pentium dummies, rm xed e ects), and unobserved (to the econometrician) characteristics, respectively. Income is represented by y it ; ijt is a mean zero stochastic term which is assumed to be i.i.d. across products and consumers; and it are individual speci c components. The individual speci c components are random coe cients, it = + D it + i ; i N(0; I k ) where mean preferences for observable product attributes are captured by, the matrix of coe cients, ; measures how tastes vary with these attributes and is a scaling matrix. Characteristics not observed by the econometrician that may in uence tastes are captured by the i : Consumers may decide not to purchase any of the products. The indirect utility provided from purchasing the outside option is u i0t = ln(y it ) + 0t + i0t ; where the price is normalized to zero. In industries where introductions of new products are frequent (like the PC industry), the assumption that consumers are aware of all products for sale is not an innocuous one. Goeree (2008) develops and estimates a model of limited information, and this study uses the same model and parameter estimates to conduct counterfactual merger analysis. In the limited information framework, the probability consumer i purchases a product depends upon the probability she is aware of product j, the probability she aware of the other products competing with j; and the probability she would buy product j given her choice set. Speci cally, let C j be the set of all possible choice sets that include product j. Assuming consumers are aware of the outside option with probability one, the (conditional) probability that consumer i purchases product j is given by s ijt = X Y Y (y it p jt ) expfx 0 j it + jt g ilt (1 ikt ) y it + P S2C j r2s (y it p rt ) expfx 0 r it + rt g l2s k =2S where ijt is the (estimated) probability consumer i is informed about product j, the outside sum is over the di erent choice sets that include product j, and the y it in the denominator 8 (2)

10 is from the outside option. order to obtain simple expressions for choice probabilities. I assume the are distributed i.i.d. type I extreme value in Advertising may impact demand through the information technology function, ijt. Suppressing time notation, the advertising technology for product j for consumer i is given by ij = exp j + ij 1 + exp j + ij (3) The components of ij that are the same for all consumers is given by j = 4 P m=1 a jm (' m + m a jm + f) + #x age j : The j term is a function of medium advertising (a jm ) which may have di erent informational e ects across media (as measured by the parameters ' m and m ) and across rms (as measured by rm xed e ects f). 10 Finally, consumers may be more likely to know a product the longer it has been on the market captured by # where x age j measured in quarters. The ij captures consumer information heterogeneity: ij = 4 P m=1 a jm ( m D s i + im ) + e D 0 i e ln i N(0; I m ): is the age of the PC Medium advertising may have di erent e ects across consumers, where Di s is vector of observed consumer attributes from the Simmons data. 11 The m Di s term is the exposure of individual i to medium m; and the term measures how advertising exposure (measured by a jm m D s it) impacts the information technology. Consumers information may be di erent even if they haven t seen an advertisement, which may depend upon their own attributes (where the e D i term is a subset of consumer characteristics). The i vector are unobserved (to the econometrician) consumer heterogeneity with regard to ad medium e ectiveness. I assume are independent of other unobservables. 12 I assume the consumer purchases at most one good per period, that which provides the 10 I include xed e ects for those rms that o ered a product every quarter, but do not estimate a separate xed e ect for each medium. 11 Simmons data are used to identify : 12 Notice ij depends upon own product advertising only. I assume the probability a consumer is informed about a product is (conditional on her attributes) independent of the probability she is informed about any other product. Information provided (through advertising) for one product (or by one rm) cannot spillover to another product (or to another rm). That is, I assume product or group advertising for product r 6= j provides no information about j. 9

11 highest utility from all the goods in her choice set. 13 of individual characteristics. Let { i = (y i ; D i ; i ; i ) be the vector The set of variables that results in the purchase of good j is R j f{ : U({; p j ; x j ; a j ; j ; ij ) U({; p r ; x r ; a r ; r ; ir ) 8r 6= jg: The market share of product j is Z Z s jt = df (y; D; ; ; ) = R j s ij df y;d (y; D)dF ()df () R j (4) where F () denotes the respective known distribution functions. The second equation follows from independence assumptions Firm Behavior I assume there are F rms in an oligopolistically competitive industry and that they are noncooperative, Bertrand-Nash competitors. J f. Suppressing time notation, the pro ts of rm f are X X (p j mc j )Ms j (p; a) + j2j f Each rm produces a subset of the J products, j2j f nh j (p nh ) X m mc ad jm( X j2j f a jm ) C f (5) where s j is the vector of home market shares, which is a function prices and advertising for all products; mc j is the marginal cost of production; nh j is the gross pro t (before advertising) from sales to the non-home sectors; mc ad jm is the marginal cost of advertising in medium m; a jm is the number of medium m advertisements; and C f are xed costs of production. The potential market size, M, is given by the number of US households in a given period, as reported by the Census Bureau. Following BLP, I assume mc j are composed of unobserved (! j ) and observed (w j ) cost characteristics: ln(mc j ) = w 0 j +! j : (6) 13 This assumption may be questionable in markets where multiple purchase is common. However, it seems reasonable to restrict a consumer to purchase one computer per quarter. Hendel (1999) presents a multiple-choice model of PC purchases in the business sector. 14 The market shares must be simulated. As in BLP, the distribution of consumer demographics is an empirical one. As a result there is no analytical solution for predicted shares. Furthermore consumers may not know all products for sale, but I don t observe the choice set facing any one consumer. I implement the solution provided in Goeree (2008) and simulate the choice set facing an individual. This eases computation burden signi cantly. Instead of requiring 2 J 1 purchase probability calculations for each individual (corresponding to each possible choice set) for each product, I need only compute the purchase probability for i corresponding to i s simulated choice set for each product. See Goeree(2008) for details on the simulation technique. 10

12 Similarly, I assume mc ad jm are composed of observed components, w ad jm (such as the average price of an advertisement), 15 and unobserved components, j : ln(mc ad jm) = w ad0 jm + j j N(0; I m ): (7) I set the variance of j to one for all media channels. Given their products and the advertising, prices, and attributes of competing products, rms choose prices and advertising media levels simultaneously to maximize pro ts. Firms may sell to non-home sectors (such as the business, education, and government sectors). Constant marginal costs imply pricing decisions are independent across sectors. 16 product sold in the home market sector will have prices that satisfy In vector form, the J rst-order conditions are Any s j (p; a) + X r2j f (p r mc r r(p; j = 0 (8) s (p mc) = 0 where j I j;r with I j;r an indicator function equal to one when j and r are produced by the same rm and zero otherwise. These FOC s imply marginal costs given by mc = p 1 s (9) However, an advertisement intended to reach a home consumer may impact sales in nonhome sectors. Optimal advertising choices must equate the marginal revenue of an additional advertisement in all sectors with the marginal cost. a jm satisfy M X r2j f (p r mc r r(p; jm Optimal advertising medium choices + mrj nh (p nh ) = mc ad jm (10) where mr nh is the marginal revenue of advertising in non-home market sectors The advertising data include ad expenditures across ten media. The quarterly average ad price in media group m is a weighted average of ad prices in the original categories comprising the group m. The weights are rm speci c and are determined by the distribution of the rms advertising across the original media. 16 There are reasons to believe that pricing decisions may not be independent across sectors. For instance, if the price of a particular laptop is lower for business, a consumer might buy the laptop from their business account for use at home. Identi cation of a model which includes pricing decisions across all sectors would require richer data for non-home sectors. Also, education, business, and government groups usually purchase multiple PCs, which greatly complicates the model (Hendel, 1999). While the assumptions that I impose imply independent pricing decisions, the estimates are sensible, and goodness-of- t tests suggest the model ts the data reasonably well. 17 I approximate the marginal revenue of advertising with mrj nh = nh p p nh j + x nh0 j nh x, where characteristics of product j sold in the non-home sector are price (p nh j ) and other observable characteristics (x nh j ) including advertising, CPU speed, and non-pc rm sales. 11

13 4 Estimation Identi cation Following the literature, I assume that the demand and pricing unobservables (evaluated at the true parameter values, 0 ) are mean independent of a set of exogenous instruments; z : E j ( 0 ) j z = E [! j ( 0 ) j z] = 0: (11) I do not observe j or! j, but market participants do. This leads to endogeneity problems because prices and ad choices are most likely functions of unobserved characteristics. A solution to this problem involves instrumental variables. 18 BLP show that variables that shift markups are valid instruments for price in di erentiated products models. In a limited information framework the components of z include the characteristics of all the products marketed (the x vector), variables that determine production costs (the components of the observed w vector that are not in x) and variables that determine advertising costs (the components of w ad ). The intuition to motive the advertising instruments follows Goeree(2008) and is similar to that used by BLP to motivate the price instruments. Products which face more competition (due to many rivals o ering similar products) will tend to have lower markups relative to more di erentiated products. Advertising for j depends on j s markup. As ad rst order conditions (FOC) in (10) indicate, a rm will advertise a product more the more they make on the sale of the product, ceteris paribus. The pricing FOCs in (8) show the optimal price (and hence markup) for j depends upon characteristics of all of the products o ered. Therefore, the optimal price and advertising depends upon the characteristics, prices, and advertising of all products o ered. Note also that the level of advertising for j in media m depends on the marginal cost of advertising in that media: Thus the instruments will be functions of attributes, product cost shifters, and advertising cost shifters of all other products. I use an approximation to the optimal instruments presented in BLP(1999). I form the approximation by evaluating the derivatives at the expected value of the unobservables ( =! = 0). The approximations are constructed so as to be highly correlated with the relevant variables and are functions of exogenous data. Hence the exogenous instruments will be consistent estimates of the optimal instruments. Details are provided in Appendix A. 18 Berry (1994) was the rst to discuss the implementation of instrumental variables methods to correct for endogeneity between unobserved characteristics and prices. BLP provide an estimation technique. My model and estimation strategy is in this spirit but is adapted to correct for advertising endogeneity. 12

14 Next, I present an informal discussion of how variation in the data identi es the parameters. The mean utility associated with purchasing a PC is chosen such that observed market shares matches predicted market shares. If consumers were identical, then all variation in sales would be driven by variation in product attributes. Variation in product market shares corresponding to variation in the observable attributes of those products (such as CPU speed) is used to identify the parameters of mean utility (). Identi cation of the taste distribution parameters (; ) relies on information on how consumers substitute. Variation in sales patterns over time as the set of available products change allows for identi cation of. Also, I augment the market level data with micro data on rm choice. The extra information in the micro data allows variation in choices to mirror variation in tastes for product attributes. Correlation between x j D i and choices identi es the parameters. If consumers were identical, then all variation in the information technology, and induced variation in shares, would be driven by variation in advertising or the age of the PC. Variation in sales corresponding to variation in PC age identi es #. Variation in sales corresponding to variation in advertising identi es the other parameters of j. Returns to scale in media advertising ( m ) are identi ed by covariation in sales with the second derivative of a jm. 19 Identi cation of rm- xed e ects ( f) are identi ed in by (i) the total variation in sales of all products sold by the rm corresponding to variation in rm advertising and (ii) in the Simmons data by observed variation in rm sales patterns corresponding to variation in rm advertising. One major drawback of aggregate ad data is that I don t observe variation across households in advertising exposure. Normally observed variation in market shares corresponding to variation in household ad media exposure would be necessary to identify and &. In Goeree (2008) I show how to use the additional information in the Simmons data to aid identi cation. In short, variation in choices of media exposure corresponding to variation in observable consumer characteristics (Di s ) identi es. Variation in sales and ad exposure (a 0 jdi s ) identi es the e ect of ad exposure on the information set (&): The other parameters of ij which do not interact with advertising ( ) e are separately identi ed from due to nonlinearities There is not enough variation in the ad data to estimate ' and e ects for all media separately. I estimate these parameters for the tv medium and for the combination of newspaper and magazine media. 20 The parameters on group advertising ( 1 and 2 ) are identi ed by observed variation in expenditures on group advertisements (ad m ) with the number of products in the group and by functional form. 13

15 Estimation Technique The estimation technique follows BLP (1995, 2004), Nevo (2000), and Petrin (2002). I use GMM to nd the parameter values that minimize the objective function, 0 ZA 1 Z 0 ; where A is a weighting matrix, which is a consistent estimate of E[Z 0 0 Z] and Z are instruments orthogonal to the composite error term. The composite error term is constructed of ve components. The rst two components are analogous to those in BLP. They arise from the demand side and the rm s optimal pricing decisions. The third component of the composite error term arises from rm s optimal advertising decisions across media. The nal two components of the composite error term use the individual level data from Simmons which provides information on individuals decisions of rm choice and individuals exposure to media. I provide more detail of the estimation technique in Appendix A. Parameter Estimates I use the estimates from Goeree (2008) to conduct the merger counterfactuals discussed in detail in the next sections. It is important to note that the estimates from the limited information model indicate that advertising signi cantly impacts the information set. Furthermore, advertising has very di erent systematic e ects across individuals. Table 3 presents parameter estimates that measure how media exposure varies with observed demographic characteristics (). These coe cients proxy for e ectiveness of ads in reaching consumers through various media. The results indicate magazines are most e ective at reaching high income individuals where the e ectiveness is increasing in household size. Newspapers are most e ective at reaching high income, married individuals who are above the age of 30. Although newspaper advertising is less likely to reach a family the larger is their household ( 0:04). TV advertising is the most e ective medium for reaching low income households. Television advertising is also e ective at reaching married individuals over 50, although not as e ective as newspaper. These results indicate that variation in ad media exposure across households is an important source of consumer heterogeneity. 14

16 Media Magazine Newspaper Television Radio Variable Coefficient Std. Error Coefficient Std. Error Coefficient Std. Error Coefficient Std. Error constant ** (0.040) ** (0.040) ** (0.041) ** (0.043) 30to50 (=1 if 30<age< * (0.025) ** (0.025) (0.025) * (0.025) 50plus (=1 if age>50) (0.025) ** (0.025) ** (0.025) ** (0.025) married (=1 if married) * (0.018) ** (0.018) ** (0.018) (0.018) hh size (household size) ** (0.006) ** (0.006) ** (0.006) * (0.006) inclow (=1 if income<$60,000) ** (0.021) ** (0.021) ** (0.021) ** (0.022) inchigh (=1 if income>$100,000) ** (0.029) ** (0.028) (0.030) ** (0.030) malewh (=1 if male and white) ** (0.018) (0.018) * (0.018) (0.018) eduhs (=1 if highest edu 12 years) ** (0.026) ** (0.026) ** (0.027) ** (0.027) eduad (=1 if highest edu 1 3 college) * (0.028) ** (0.027) ** (0.028) ** (0.029) edubs (=1 if highest edu college grad) (0.025) ** (0.024) ** (0.025) ** (0.026) edusp (education if <11) ** (0.003) ** (0.003) ** (0.003) ** (0.003) Notes: Estimates include time dummies. ** indicates significant at the 5% level; * significant at the 10% level. Table 3: Media Exposure Parameter Estimates Table 4 shows that the variation in ad exposure translates into variation in information sets with a positive and highly signi cant estimate for & (0.948). Parameter estimates of e suggest other means of information provision, such as word-of-mouth or experience, play a role in informing certain types of consumers. The coe cient on income less than $60,000 (0:69) indicates these individuals are likely to be informed about 41% of the products without seeing an ad. Whereas having a high income is not signi cantly di erent from having a middle income, in terms of being informed without seeing an ad. In addition, the probability of being informed without seeing any advertising is higher for high-school grads relative to non-graduates. Consumers are signi cantly more likely to know a PC the longer it has been on the market (0:16): There are decreasing returns to advertising in the TV ( 0:05) and newspapers and magazines ( 0:01) media, but they are decreasing at a faster rate for TV. Estimates of rm xed e ects interacted with total advertising ( ) indicate that some rms are more e ective at informing consumers through advertising. Most notably ads by Compaq, Dell, Gateway, IBM and Packard Bell are signi cantly more e ective, which could be due to di erences in advertising techniques across rms I present the parameter estimates for the remaining parameters of the utility and cost functions in Appendix B. 15

17 Variable Coefficient Std. Error consumer demographics constant ** (0.004) high school graduate ** (0.028) income < $60, ** (0.009) income > $100, (0.318) age of pc ** (0.005) interactions with product advertising media exposure * advertising ** (0.059) media advertising npand mag advertising * (0.488) tv advertising ** (0.418) (np and mag advertising) (0.014) (tv advertising) ** (0.004) firm total advertising acer (0.042) apple (0.790) compaq ** (0.077) dell * (0.460) gateway ** (0.065) hewlett packard (11.750) ibm ** (0.184) micron (5.832) packard bell * (0.149) Notes: ** indicates t stat > 2; * indicates t stat >1. Unless units are specified variable is a dummy. Table 4: Other Information Technology Parameter Estimates The estimated parameters have important implications for pricing and advertising behavior and markups. behavior of consumers. The markups earned by rms are determined, in part, by the substitution Substitution could be induced by changes in prices or choice sets, the latter of which is signi cantly impacted by advertising with varying e ects across consumers. When advertising changes the impact on the choice set is more pronounced for those consumers who are more sensitive to advertising. The rms decisions of what prices to charge and how much information to provide through advertising depend upon the price and advertising elasticities of demand. The top panel of Table 5 presents a sample from 1998 of own- and cross-price elasticities of demand. 22 The results show that products are more sensitive to changes in prices of 22 Elasticities are computed by multiplying the numerical derivative of estimated demand by price and dividing by actual sales. 16

18 computers with similar characteristics. Among PC s that have a windows operating system, form factor plays a strong role in substitution patterns. For example, Compaq Armada laptop is most sensitive to changes in prices of other laptops rather than to changes in other Compaq non-laptop computers. While Apple computers are most sensitive to changes in the prices of other Apple computers implying there is less substitution across platforms. Apple Apple Compaq Compaq Dell HP HP IBM IBM PowerBook* Power Mac Armada* Presario Latitude* Omnibook* Pavilion PC Thinkpad* price elasticities PowerBook* Power Mac Armada 7xxx* Presario 2xxx Latitude XPI* Omnibook 4xxx* Pavilion 6xxx PC 3xxx Thinkpad 7xxx* advertising semi elasticities PowerBook* Power Mac Armada 7xxx* Presario 2xxx Latitude XPI* Omnibook 4xxx* Pavilion 6xxx PC 3xxx Thinkpad 7xxx* Notes: A * indicates a laptop. For price elasticities, cell entries i,j where i,indexes row and j column, give the percentage change in market share of brand I with a 1% change in the price of j. Each entry represents the median of the elasticities from For advertising elasticities, cell entries i,j give the percent change in the market share of i with a $1000 increase in the advertising of j. Table 5: A Sample from 1998 of Estimated Price and Advertising Elasticities Estimated advertising demand elasticities indicate that, for some rms, advertising for one product has negative e ects on other products sold by that rm but it is less negative than for some of the rival products. 23 The lower panel presents a sample from Each semi-elasticity gives the percentage change in the market share of the row computer associated with a $1000 increase in the (estimated) advertising of the column computer. For instance, a $1000 increase in advertising for Apple Power Mac results in a decreased market share of around 0.1% for Compaq Presario but has very little e ect on the market share for Apple PowerBook. In contrast, an increase in advertising for HP Omnibook has a large e ect (relative to increase in own market share) on the market share for HP Pavilion. 23 The model does not allow advertising for one product (or by one rm) to have positive spillovers to another product. Hence, the cross-product advertising e ects (the o -diagonals in the lower panel of Table 5) are all negative. The diagonal elements report the increase in market share from own-advertising. For example, an increase of $1000 for advertising on Dell Latitude results in an increased market share of 0.02%. 17

19 Median Percentage Markup Average Price Ad to Sales Ratio over marginal costs including ad costs Total Industry $2, % 15% 10% Top 6 firms $2, % 17% 12% Apple $1, % 16% 9% Compaq $2, % 23% 16% Gateway $2, % 12% 10% Hewlett Packard $2, % 16% 10% IBM $2, % 16% 10% Packard Bell NEC $2, % 16% 11% Notes: ad to sales ratios are annual averages and are from LNA and include all sectors (home, business, education, government). markups are the median (price marginal costs)/price across all products. The last column is determined from estimated markups and estimated effective product advertising in the home sector. Table 6: Estimated Median Markups I use estimated demand to infer marginal costs and markups as shown in Table 6. median markup charged by PC rms is 15% over marginal costs of production and 10% over per unit production and (estimated) advertising costs. As the rst two rows show, the top rms have higher than average markups and advertising expenditures relative to the industry. Indeed the non-top rms average median markup is much lower, 12%, with an ad-to-sales ratio of about 2%. The The nal column shows that, even after controlling for the fact that the top rms advertise more, they continue to earn higher than average markups. In 1998 the median industry markup was 19% over costs with the top rms earning a 22% markup. Overall industry and top rm markups were increasing over the period. 5 Post Merger Equilibrium In this section I present the results from two counterfactual merger simulations: a merger between HP and Compaq and a hypothetical merger between IBM and Dell. All post-merger computations use the estimates from the limited information model of demand and supply presented in the above section. I simulate post-merger equilibria under two assumptions of post-merger behavior on part of the rms. First, I compute the new price equilibrium under the assumption that advertising choices remain at pre-merger levels. That is, I do not allow rms to reoptimize over advertising choices. This is used as the benchmark case. Notice that this benchmark case will provide an accurate picture of the post-merger industry only if rms do not change their advertising strategy or if advertising does not impact demand. 18

20 I compare the benchmark case to the new price and advertising equilibrium that arises post-merger allowing rms to reoptimize over both prices and advertising. For the rst two parts of the analysis I use estimated costs and, hence, assume the cost structure is the same both before and after the merger. Firms may experience decreased costs (synergies) as a result of a merger. Therefore, in the nal part of the analysis (following Nevo, 2000) I determine the magnitude of cost savings that would be necessary to return to the pre-merger price and advertising equilibrium. I assume post-merger market conduct is the same as pre-merger, and use the estimated pre-merger parameters to simulate the (counterfactual) post-merger equilibrium. The predicted post-merger equilibrium price p post solves p post = cmc + post (p post ) 1 s(p post ) (12) where the matrix post is constructed to re ect post-merger ownership structure, cmc are the predicted marginal costs and p post is the vector of post-merger predicted equilibrium prices. The predicted post-merger advertising levels solve (10) under the new ownership structure, holding marginal costs constant at their estimated levels. The post-merger equilibrium levels of prices and advertising are simulated jointly using the data from the last quarter of Advertising and Market Power Table 7 presents rm and industry level changes in prices, advertising, and markups after the mergers. The rst column presents average percentage changes in prices under the assumption that post-merger advertising levels are unchanged. Predicted price changes are higher for the merging rms, especially under the Compaq-HP merger. In addition all rms increase their prices under the HP-Compaq merger. We might expect prices to increase for all rms in the industry, as the industry is more concentrated allowing rms to exercise more market power. However, counter to intuition, under the Dell-IBM merger some rms charge lower average prices and overall industry prices experience only a small increase in price (0.2%). This unexpected result may be due to the assumption that rms leave advertising expenditures unchanged. 19

21 Optimizing Variable: Price only Price and advertising % change % change Difference in % change Estimated in prices in prices price changes in ads % Markups Compaq HP merger: Apple 2.1% 5.8% 3.6% 0.3% 12% Compaq HP 18.9% 18.4% 0.5% 10.8% 31% Dell 2.9% 2.9% 0.0% 3.3% 13% IBM 5.8% 7.8% 2.0% 0.2% 25% Industry Total 6.9% 9.3% 2.4% 3.3% 24% Dell IBM merger: Apple 0.0% 4.8% 4.8% 0.2% 13% Compaq 0.1% 7.8% 7.8% 2.4% 24% Dell IBM 1.3% 5.9% 4.6% 0.9% 13% Hewlett Packard 0.5% 4.8% 4.3% 8.5% 20% Industry Total 0.2% 6.1% 5.9% 1.8% 21% Note: Each entry represents the average percentage change in the final quarter of Estimated % Markups include ad costs Table 7: Firm and Industry Post-Merger Results The second and third columns present results based on the assumption that rms choose new prices and advertising levels after the merger. As the second column shows, the pricing outcome is more intuitive. All rms increase their prices as industry concentration grows. The overall price increase in the industry is 9% and 6% under the HP-Compaq and Dell-IBM mergers, respectively. Surprisingly, Dell-IBM do not raise prices as much as Compaq under the Dell-IBM merger. The rst column of Table 8 provides a breakdown of price changes for selected products. As the table indicates, price increases are greatest for the HP and IBM products under their respective mergers. One possible reason for this distribution is that both HP and IBM have lower pre-merger prices than their merging counterparts. As the third column of Table 7 indicates, prices are 2% and 6% higher under the HP-Compaq and Dell-IBM mergers respectively relative to when rms don t reoptimize over advertising. The fourth column presents percentage changes in advertising. All rms choose to advertise more in both post-merger environments. Advertising increases by 3.3% under the HP-Compaq merger and 1.8% under the Dell-IBM merger. Prior to merging HP-Compaq have a combined ad-to-sales ratio of 7%, while Dell-IBM s is on the order of 14%. HP- Compaq increases their advertising by 10% (or $75 million) after the merger, yielding a post-merger ad-to-sales ratio of 20%. 24 Dell-IBM increases their advertising by 0.9% (or $12 million) post-merger, yielding a post-merger ad-to-sales ratio of 22%. It is interesting that post-merger ad-to-sales ratios are similar for the merging rms. 24 This is calculated using the post-merger shares, which are lower than pre-merger combined shares. 20