Advertising and quality in the U.S. market for automobiles.

Nichols, Mark W.

1. Introduction

Each year, automobile manufacturers spend millions of dollars on advertising. Between January and September 1995, for example, General Motors (GM) spent $1.07 billion (Crain Communications, Inc. 1996). What function do such large outlays serve, and what is the correlation between these outlays and a vehicle's quality? Do high-quality producers use advertising as a means to signal their quality to relatively uninformed consumers or do low-quality producers take advantage of this information asymmetry and mimic their high-quality competitors, thereby inhibiting advertising's use as a signal? On the other hand, does advertising merely serve an informational role, reminding consumers that a product's recent ancestors proved to be high quality?

These questions continue to be a perennial debate in economics. While recent game-theory literature thoroughly analyzes advertising's ability to signal quality (reviewed below), empirical evidence is somewhat limited and generally mixed. For example, Caves (1986) and Tellis and Fornell (1988), using the Profit Impact of Market Strategies database (a broad interindustry data set), show that high quality generally induces higher advertising expenditures. In contrast, Caves and Greene (1996), conducting a multiple-product analysis of quality, price, and advertising, find a positive correlation between advertising and quality when examining goods where buyers' experience and search are effective at guiding brand choice but a negative correlation for convenience goods.

One explanation for these mixed findings is the use of broad cross-sectional data sets and treatment of advertising as a homogenous activity in different markets. Indeed, the findings of Caves and Greene (1996) support the idea that the use of advertising varies across industries depending on market structure, product characteristics, and consumer characteristics and that "empirical studies using individual industries as cross-sectional observations may be economically uninterpretable" (Leffler 1981, p. 46).

In an attempt to avoid this potential pitfall, the present study employs data from the U.S. automobile market and compares outlays on superior products relative to inferior products for producers of otherwise identical goods, that is, it conducts an intraindustry analysis in order to avoid any potential bias introduced by analyzing multiple-product, cross-sectional data. Nevertheless, even in an intraindustry study there are several confounding factors to consider. For example, consumers' knowledge of the product will influence advertising. Ceteris paribus, greater knowledge about product quality should result in lower advertising expenditures. In addition, there may exist manufacturer-specific factors, such as the number of models produced, that influence advertising. A detailed description of the model and controlling factors employed in this study are provided in Section 3.

The automobile industry is chosen to conduct the intraindustry analysis because model-specific information on advertising outlays and product quality is readily available. More importantly, however, the quality ratings for automobiles employed in this study are unobservable when a car is new and being advertised. The ratings, which first become available a year after a vehicle's introduction, therefore reflect information on quality that is unknown to the consumer at the time of purchase. This feature, current advertising expenditures and unknown quality, provides a unique opportunity to empirically examine hypotheses offered by the recent game-theory literature on the ability of advertising to signal product quality to relatively uninformed consumers.

Before proceeding, it should be noted that, while quality ratings on the current model being advertised are unavailable, ratings on a previous model's quality do exist. To the extent that quality ratings are highly correlated over time, past quality may be a reliable indicator of current quality. In short, current quality may not be completely unknown to consumers. If true, advertising's role may be to simply inform or remind consumers of high quality in the past. To distinguish between these two hypotheses (signalling versus information), the current study controls for past quality, examines how changes in the (future) quality rating from past quality ratings impact advertising, and compares advertising and quality for introductory models where no past data exist. The results reveal that advertising is multidimensional and serves both as a signal of quality and a provider of information. Nevertheless, strong support for the signalling hypothesis is found when controlling for past quality and when examining changes in quality. Nowhere in the analysis is it found that low-quality automobiles are advertised more intensively.

The next section reviews theories on the strategic use and determinants of advertising. Section 3 describes the empirical model and the choice of variables to empirically test the advertising-as-a-signal hypothesis. This is followed by a discussion of the methodology used to derive the results reported in section 4. Section 5 concludes the paper.

2. Theories on the Strategic Use and Determinants of Advertising

There is a vast theoretical literature addressing the strategic use of advertising and its correlation to quality. Much of this literature focuses on experience goods whose quality is costly to ascertain prior to purchase and is only learned over time with consumption (Nelson 1970). While this literature varies dramatically in its assumptions and conclusions, a central theme is advertising's ability to signal quality to uninformed consumers.

The origin of the signalling literature can be traced back to Nelson (1970, 1974). In this series of seminal articles, Nelson distinguishes between search and experience goods and the informational content of advertising. In the case of search goods, advertising provides direct, credible information about product characteristics and quality because consumers can verify this information prior to purchase. With experience goods, however, quality is not verifiable prior to purchase. Consequently, claims of high quality lack credibility because they can be freely made by all producers. Realizing this, consumers will rationally ignore any direct claims about high quality. Nevertheless, the consumer can extract indirect information. In particular, the consumer learns that the brand is being advertised. Therefore, in the case of experience goods, it is the level of advertising, not necessarily the information content, that provides information to consumers.

The mechanism leading to a positive correlation between advertising levels and quality in Nelson's model is repeat purchases. There are two ways that repeat purchases induce a positive correlation between advertising and quality. First, when only high-quality products induce repeat purchases, high-quality producers have a greater incentive to advertise in order to increase demand and expected profits. Second, because high-quality products induce repeat purchases, high-quality producers will wish to distinguish themselves from low-quality producers. For this to occur through advertising, low-quality producers must be unable to recoup the costs of advertising necessary to mimic the advertising strategies of high-quality producers.

There are subtle, but distinct, differences between these two strategies. In the first scenario, firms are not intentionally signalling their quality. Instead, they are merely trying to increase demand in the initial period in order to induce future sales and increase expected profits. In the second scenario, however, the firm is intentionally signalling its quality to consumers, thereby distinguishing itself from its low-quality competitors. In this latter case, it is immaterial whether advertising has any direct impact on demand.

More recent studies on advertising as a signal of quality have followed the second scenario, focusing on cost differences between high- and low-quality firms. For example, Khilstrom and Riordan (1984) provide a formal game theoretic model to Nelson's argument where higher fixed, but not marginal, costs for high-quality firms result in advertising signalling quality. This requires that high-quality firms be able to recover their advertising costs (through repeat sales) while low-quality firms cannot and that consumers know enough about cost differences to realize that only high-quality firms can profitably advertise.

Klein and Leffler (1981) model advertising as a signal of the existence of a firm-specific selling cost and the existence of a price premium. This, too, results in a positive correlation between advertising and quality. In particular, knowing that sunk costs (advertising) are only profitable if the expected future quasi-rents exceed the one-time gain from cheating, that is, providing low quality at the higher price, larger advertising expenditures signal larger price premiums and higher quality.

Finally, Milgrom and Roberts (1986) extend this literature by providing a model where a monopolist signals quality through both price and advertising or other dissipative expenditures, The extension is perhaps best understood in the context of Klein and Leffler, recognizing that the price premium enjoyed by high-quality firms can be used by consumers to infer quality.

All of the above studies are valuable contributions to the signalling literature. However, the ability to directly test them empirically is limited. Very little of the signalling literature deals with an oligopoly setting where the actions of rival firms may have an impact on advertising decisions. In fact, Milgrom and Roberts (1986, p. 802) note that "the assumption of monopoly seems natural in this context, at least in comparison with the perfectly competitive alternative. Treating the intermediate case of oligopoly would involve significant additional problems." Nevertheless, there is no reason to rule out the possibility that the actions of rivals are an important determinant of a firm's advertising decision in an oligopolistic market such as automobiles (see, e.g., Kwoka [1993], who shows that the action of rivals influences a firm's sales).

Another factor clouding any empirical test is the fact that most of the signalling literature assumes advertising is purely dissipative and has no direct effect on demand.(1) This seems implausible given that advertising for automobiles, in addition to being a potential signal of quality, will carry information about price, warranty, style, etc. Firms may be just as likely to advertise to increase demand as to signal quality. Therefore, there is a need to consider other factors that influence the firm's advertising decision and to consult the non-game-theoretic models that address the determinants of nondissipative advertising.

One of the first papers to examine the optimal advertising intensity was Dorfman and Steiner (1954). Specifically, for a profit maximizing firm, the advertising-to-sales ratio will equal the ratio of the advertising demand elasticity to the product demand elasticity or, similarly, the product of the price cost margin and the advertising elasticity. Conclusions arising from these optimality conditions are that products with lower demand elasticities and/or higher profit margins will have higher advertising expenditures.

More recent non-game-theoretic studies have examined the determinants of advertising using a derived demand framework (e.g., Porter 1976; Ehrlich and Fisher 1982; Caves 1986). In these models, as with signalling models, consumers have imperfect knowledge about product quality. Therefore, consumers seek information from various sources such as word of mouth, technical publications, and advertising. It is this demand for information that leads to a derived demand for advertising. According to Ehrlich and Fisher (1986, p. 366), "[Advertising] can be constructed on the assumption that it affects not tastes, but the opportunities under which advertised products can be acquired. [Advertising] affects the demand for goods because it lowers the gap between the market price received by the seller and the full price borne by the buyer." While consumer knowledge about a product is clearly an unknown, variables capturing the newness of the product are used as proxies (Caves 1986).

The non-game-theoretic approaches to advertising are useful for examining the determinants of advertising but unfortunately, unlike the game-theoretic signalling literature, depicting the relationship between advertising and quality is not their primary purpose. Moreover, they often will not examine why consumers respond to advertising. Rather, consumers are simply assumed to respond to advertising, many times according to a specific or ad hoc functional form (e.g., Schmalensee 1978).

In summary, both the game- and non-game-theoretic models make varying assumptions and offer useful conclusions about the levels of advertising. However, neither is completely satisfactory for the purpose of empirically testing the relationship between advertising and quality. The signalling literature generally ignores the impact of advertising on demand, while the non-game-theoretic literature provides little theoretical basis for the correlation between advertising and quality. Nevertheless, it seems plausible that advertising could be used by firms to both signal quality and influence demand. Unfortunately, firms do not separate their advertising expenditures according to these categories. Instead, only total advertising is observed. Consequently, the empirical model chosen for this study adopts ideas from both sets of literature and examines how advertising varies with quality and various other determinants of advertising.

Before proceeding with the empirical model, it should be noted that other scholars have extended the advertising/signalling literature by allowing for multiple signals such as advertising and price (Milgrom and Roberts 1986) and price and warranties (Lutz 1989). In the matter of warranties, Lutz offers a model where warranties serve as a signal of quality when high-quality products are cheaper to warrant than low-quality products.

A complete empirical analysis of all potential signals of quality is a daunting task, and this paper is merely a first attempt at analyzing a very difficult problem. Consequently, the analysis focuses primarily on the advertising/quality link. Warranties are difficult to analyze because there is little variation between brands. Moreover, there is evidence that interpreting warranties is costly to consumers, limiting their ability to signal quality (Crocker 1986). Similarly, price as a signal of quality is difficult to analyze because producers may choose not to simultaneously use price and advertising to signal quality when signal loss is possible (Hertzendorf 1993). When consumers cannot distinguish between no advertising and not seeing the advertisement (TV not turned on), low-quality producers can successfully mimic the pricing strategies of high-quality producers and price will not vary with quality.

3. Empirical Model and Data

To examine the relationship between the levels of advertising and quality and other various determinants of advertising, the following equation is estimated:

[ADV.sub.j,t] = [[Beta].sub.0] + [[Beta].sub.1][q.sub.j,t] + [Gamma][K.sub.j,t] + [Pi][C.sub.j,t] + [Lambda][X.sub.j,t] + [[Epsilon].sub.j,t], (1)

where [ADV.sub.j,t] represents advertising for model j at time t, [q.sub.j,t] represents the quality of model j at time t, [K.sub.j,t] proxies for buyer knowledge through variables that capture the newness of model j at time t, [C.sub.j,t] represents industry and manufacture specific conditions that affect advertising, [X.sub.j,t] represents other control variables that affect advertising, and [Gamma], [Pi], and [Lambda], are vectors of coefficients.

To estimate Equation 1, this study employs a dataset consisting of various car models (e.g., Ford Thunderbird, Toyota Corolla) produced by U.S. and foreign manufacturers for the model years 1985-1990. Only models where consistent trouble indices were reported by Consumer Reports (CR) (Consumers' Union 1986-1995) are included. Over 100 models are available in each model year, yielding a sample of 630 observations. Nominal dollar values are adjusted for inflation using the consumer price index (CPI). A description of the variables included in the empirical analysis and their mean values are provided in Appendix A. The choice of particular variables and their expected impact on advertising expenditures is discussed below.

Advertising expenditures [ADV.sub.j,t] are drawn from Leading National Advertisers Class/Brand Year to Date for 1984-1990. Expenditures are reported for different categories, including magazines, newspaper supplements, network television, spot television, network radio, and outdoor. Total advertising is defined as the sum of these categories. Television advertising is the sum of network and spot television, while print advertising includes magazine and newspaper supplements. Advertising in newspapers and cable television are not included because they were not available for the entire sample period.(2)

Quality [q.sub.j,t] is defined as the average of all available trouble indices for model j reported by CR.(3) A value of one is given if a car is much better than average, while a value of five is given to vehicles that are rated much worse than average. Thus, higher values of the trouble index reflect poorer quality.(4) Again, this information on quality is not observed by consumers when a car is new. If advertising is used to signal high quality, larger advertising expenditures should be associated with smaller values of the trouble index.

The next set of variables in Equation 1, [K.sub.j,t], pertain to the newness of the car model and proxy for consumer prior knowledge of the product. [K.sub.j,t] includes dummy variables equal to one if the model has been mechanically redesigned, if the model had significant styling changes, that is, all new body panels, and if the vehicle is an introductory model launched for the first time.

Vehicles that have been mechanically redesigned are expected to have higher advertising, ceteris paribus, since consumers may have limited knowledge about these changes or their effects on the car's performance. Styling changes, in contrast, are highly visible and there may be less demand for information from the producer. Consequently, styling changes are expected to have little impact on advertising.

Consumers have little or no information on introductory models. For example, consumers had limited information to judge the quality of the Ford Taurus when it was first introduced in 1986. Moreover, manufacturers may simply want to make consumers aware that a new model exists. Therefore, advertising levels for introductory models are expected to be higher, ceteris paribus.

Industry and manufacturer conditions [C.sub.j,t], also enter the advertising equation. The number of models offered by a manufacturer, total sales of model j at time t, price, and engineering and styling changes of rival manufacturers are included.

The number of models offered by any particular manufacturer is expected to affect advertising, but its directional impact is ambiguous. If spillovers or externalities exist among advertised models, producers offering more models may have lower advertising expenditures. On the other hand, a larger number of models may require higher advertising outlays given the increase in consumer choice.

Ceteris paribus, higher sales are predicted to be associated with higher advertising expenditures. Dorfman and Steiner (1954) have shown that the advertising-to-sales ratio should equal the ratio of the advertising-demand elasticity to the product-demand elasticity. Thus, if elasticities are relatively constant, firms will maintain a constant advertising-to-sales ratio.(5)

Yet another application of Dorfman-Steiner is that models with higher profit margins will have higher advertising expenditures. If profit margins and quality are positively correlated, failure to account for this may bias the empirical results. Unfortunately, model-specific profit margins are unavailable. Therefore, market price for model j at time t + 1 is used as a proxy.(6) Ceteris paribus, advertising is predicted to be greater for higher priced, that is, markup, vehicles.

The actions of rival manufacturers also enter the advertising function shown in Equation 1. Rivals are defined as other manufacturers' models competing in the same class. For example, rivals to the Cadillac are all large, non-GM cars. Styling and engineering changes by rivals are expected to increase advertising outlays.

Finally, included as other control variables, [X.sub.j,t], are a trend variable and manufacturer and class-size dummies. The trend variable captures the changing expenditures on advertising over the sample period and controls for any macroconditions that simultaneously affect the advertising levels of all automobile manufacturers. Dummy variables for U.S. (GM, Ford, and Chrysler) manufacturers control for manufacturer-specific effects (foreign is the excluded category). Finally, dummies for compact, medium, and large (small is the excluded category) account for any class-specific effects. Berry, Levinsohn, and Pakes (1995) have shown that crowded segments, such as small and compact vehicles, have greater price elasticities of demand. This suggests that these vehicles will be advertised less, ceteris paribus, a la Dorfman-Steiner.

4. Methodology and Results

Before presenting results, a brief discussion of the specification of the advertising equation is in order. The advertising equation is estimated using two-stage least squares with [SALES.sub.j,t-1], [REALPCY.sub.t], and [GASPRICE.sub.t] serving to (over) identify the [ADV.sub.j,t] and [SALES.sub.j,t] relationship.(7) The suitability of these instruments is confirmed by the high partial [R.sup.2] (0.85) and F statistic (850.26) for the excluded instruments.(8) A semilog model is estimated because it provides a better fit using a Box-Cox (Box and Cox 1964) test.(9) Applying a Koyck transformation results in the following equation to be estimated:

[Mathematical Expression Omitted], (2)

where [ADV.sub.j,t-1], i = 0, 1, are log values and [q.sub.j,t], [K.sub.j,t], [C.sub.j,t], and [X.sub.j,t] are defined as above.(10)

Table 1. Average Trouble Index/Change in Average Trouble Index for
U.S., Leading Japanese and Foreign Manufacturers

Year U.S.(a) Leading Japanese(b) Other Foreign(c)

1985 3.75/- 0.16 1.33/0.01 2.41/0.19
 (0.82)/(0.45) (0.57)/(0.44) (0.95)/(0.55)
1986 3.64/-0.20 1.29/-0.01 2.34/0.02
 (0.83)/(0.48) (0.39)/(0.15) (0.72)/(0.46)
1987 3.61/-0.22 1.46/0.15 2.37/0.13
 (0.88)/(0.60) (0.66)/(0.41) (0.91)/(0.67)
1988 3.50/-0.20 1.35/0.00 2.54/0.22
 (0.88)/(0.78) (0.54)/(0.30) (1.02)/(0.80)
1989 3.60/-0.13 1.31/-0.04 2.55/0.18
 (0.96)/(0.89) (0.45)/(0.40) (1.25)/(0.87)
1990 3.38/-0.22 1.33/0.10 2.60/0.18
 (0.89)/(0.92) (0.33)/(0.46) (1.13)/(0.74)

Change in average trouble index is the current model's average
trouble index minus the average trouble index of last year's model.
Standard deviation in parentheses.

a U.S. automobiles are GM, Ford, and Chrysler.

b Leading Japanese automobiles are Honda, Nissan, and Toyota.

c Other foreign consists of BMW, Hyundai, Mazda, Mercedes Benz,
Mitsubishi, Saab, Volkswagen, and Volvo.

Table 1 reflects the pattern of quality change over the sample period by showing the level of the average trouble index and its average annual change, defined as the average trouble index of the current model minus the average trouble index of last year's model, for U.S., leading Japanese, and other foreign automobiles. Confirming general market perceptions, Japanese automobiles have higher quality than the other manufacturers. Moreover, there is less variation in quality among the various Japanese models, as indicated by the smaller standard deviation of the average trouble index. Table 1 also reveals that the relative quality of U.S. automobiles improved over the sample period. The change in the average trouble index is negative for all years, and while the overall reduction in the average trouble index may appear small, 3.75 in 1985 to 3.38 in 1990, it is important to keep in mind that this measure cannot account for improvements in all automobiles over time since CR's ratings are model-year specific. Thus, [TABULAR DATA FOR TABLE 2 OMITTED] the rising value of the average trouble index for other foreign cars does not indicate that 1990 models are of lower quality than 1985 models since the quality of all vehicles has improved over time. It does indicate, however, that relative improvements among foreign cars are occurring below the average rate. The end result is a deterioration in the relative quality of foreign automobiles, causing their average trouble index to increase. By the same reasoning, U.S. automobiles are improving at a faster rate, causing their trouble index to decline.(11) This feature, differing levels of quality and changes in quality over time, make the automobile industry, and U.S. automobiles in particular, an interesting case study to examine advertising's role in conveying information and signalling quality.

Table 2 shows the simple correlation between the average trouble index and total advertising expenditures and the average trouble index and price by model year. Generally, the results show a negative correlation, reflecting a positive relationship between quality and advertising and quality and price because higher values of the trouble index reflect poorer quality. While this does not hold in model years 1985 and 1988 for advertising and quality, the positive correlation between advertising and the trouble index in those years is generally small.

Table 3 provides evidence on real advertising expenditures per model over the period and the allocation across television and print media. Total expenditures on advertising have risen over the sample period, with the largest increase coming between 1987 and 1988. Moreover, nearly all of the increase has been due to greater outlays on television advertising. Print advertising, in contrast, rose only slightly, and by 1990 had nearly returned to its 1985 level. On average, the U.S. has lower advertising expenditures per model than foreign manufacturers.(12)

Two-stage least squares estimates of Equation 2 are provided in Table 4, with total, television, and print advertising equations reported. All variables are of the expected sign, and several are significant. The variable of most interest for our purpose is the average trouble index. The results for total advertising suggest that an above-average model (a value of two) would have advertising expenditures approximately 15% higher than an average model (a value of three). Moreover, much of the advertising is seen to occur on television, where an above-average model has expenditures approximately 19% higher. The reliance on television advertising is intriguing given its reputation for excessive expenditures on fluff or noninformative material.(13) While generally providing few hard facts, little factual information can be credibly conveyed for experience goods since much of the information is unverifiable. Television, therefore, provides an excellent medium to signal quality through conspicuous, flashy commercials, whose primary information is that the firm spent a lot of money for the advertisement. While the coefficient on the trouble index for print advertising is insignificant, the above results generally support the hypothesis that higher quality products are more heavily advertised.(14)

Table 3. Mean Advertising Expenditure per Model by Media Type(a)

Year Total Television(b) Print(c)

U.S. Automobiles(d)

1985 9502 6466 2885
1986 7556 5011 2432
1987 8623 5676 2793
1988 13,081 9341 3394
1989 11,848 8583 2900
1990 13,491 10,275 2896

Foreign Automobiles(e)

1985 8713 5428 3121
1986 9430 6250 3045
1987 11,520 7729 3650
1988 12,836 8969 3646
1989 15,691 11,281 3950
1990 15,512 11,935 3157

a Expressed in thousands of dollars.

b Television consists of network and spot coverage.

c Print includes magazines and newspaper supplements.

d U.S. automobiles are GM, Ford, and Chrysler.

e Foreign automobiles are all others.

As expected, sales are shown to positively influence advertising. An increase in sales of 1000 is associated with a 0.29% increase in advertising expenditures (approximately $32,500 using the average advertising expenditure given in Appendix A). The existence of spillover effects in advertising, where consumers associate quality with the manufacturer (e.g., Ford) rather than the model (e.g., Escort), is supported by the fact that manufactures with more models have lower expenditures per model, ceteris paribus.

While style changes have been shown by Kwoka (1993) to increase sales, they have no effect on advertising expenditures. Style changes are easily discernible by the consumer, and unlike quality and engineering changes, there is less ambiguity about their effect. Engineering changes, however, are less visible to the consumer and are likely to be associated with improved quality due to better handling or improved body design. While positive, the coefficient is not significant at conventional levels. The positive and significant coefficient for introductory models, [TABULAR DATA FOR TABLE 4 OMITTED] however, is consistent with the hypothesis that little knowledge about them exists and that information is transmitted, in part, through advertising.

Finally, the positive coefficient on price suggests that manufacturers advertise higher priced (i.e., higher profit margin) vehicles more intensively, ceteris paribus. Given the positive correlation between quality and price in Table 2, this result is also consistent with the simultaneous use of price and advertising as a signal of quality. The result for total advertising, however, is not significant at conventional levels. The coefficients on the manufacturer dummies indicate that GM advertises more than foreign manufacturers, whereas Ford and Chrysler do not. In addition, television advertising on compact vehicles is significantly higher than expenditures on small cars. Generally, however, compact, medium, and large cars have advertising expenditures that are not statistically different from small vehicles.(15)

The above results demonstrate that higher quality vehicles are more heavily advertised, even when controlling for other influential factors such as profit margins, elasticity of demand, technological changes, etc. These expenditures may not be signalling new information to consumers, however, if the quality of this year's model is highly correlated with the quality of earlier models. In short, last year's quality may serve as a sufficient indicator of this year's quality. To account for this, Equation 2 was estimated including the quality of the previous year's model.(16) The results, provided in Appendix B, are nearly identical to those reported in Table 4, supporting the advertising-as-a-signal hypothesis as opposed to the hypothesis that advertising solely provides information about past quality.(17) Nevertheless, if advertising truly signals high quality, the largest expenditures should occur when quality changes significantly from the previous year or on high-quality introductory models, where little or no previous information exists.

To test the hypothesis that advertising is higher when quality is most improved, the change in the current model year's average trouble index from the previous year is included in the advertising equation.(18) The previous model year's average trouble index is also included to control for the absolute level of quality. The results, presented in column 1 of Table 5, show that advertising expenditures significantly rise when quality improves from the previous year. The coefficient on the change in the average trouble index is negative and statistically significant. This suggests that manufacturers are signalling their quality improvements to consumers through higher advertising expenditures. Coefficient estimates on the remaining variables mirror those described above.

Column 2 of Table 5 introduces the average trouble index interacted with the introductory model dummy variable. While the coefficient on the interaction term is negative, it is not statistically significant at conventional levels. While earlier results revealed that producers are advertising both introductory models and higher-quality models more intensively, higher-quality introductory models are not advertised significantly more than older models with similar quality. A potential explanation for this result is that advertising's role of providing information dominates its use as a signal of quality in the case of introductory models.

As a final investigation into the advertising behavior of the auto manufacturers, column 3 of Table 5 provides results when the average trouble index is replaced with dummy variables representing each level of quality. To be included in a category, a model must have received a majority (at least three out of five) of those rankings. For example, a vehicle that was rated [TABULAR DATA FOR TABLE 5 OMITTED] much better than average (trouble index equal to one) for three out of five years and better than average (trouble index equal to two) for two of the five years would be considered much better than average. On the other hand, a vehicle rated much better than average for two of the five years and better than average for three would be classified as better than average. The results indicate that manufacturers do indeed advertise above-average quality cars more heavily. In contrast, advertising levels for average and worse than average cars are indistinguishable for models whose quality is much worse than average (the omitted category). This is consistent with a separating equilibrium where only the highest quality firms incur large advertising expenditures.

Finally, column 4 of Table 5 replaces [SALES.sub.j,t] with the market share of model j at time t, defined as model j's sales at time t over total retail passenger car sales at time t. Bagwell and Ramey (1993) suggest that firms incur greater advertising expenditures because they have larger market shares rather than higher quality. The results support this argument, in part. Manufacturers do incur higher advertising expenditures on models with larger market shares. However, higher quality vehicles are still found to be more heavily advertised, ceteris paribus.

5. Concluding Remarks

This paper provides an empirical investigation into advertising's role as a source of information and signal of quality in the U.S. market for automobiles. The results reveal that advertising serves both roles by providing information and signalling quality to imperfectly informed consumers. Advertising expenditures are 15% higher for an above average quality car relative to an average quality vehicle. Moreover, a majority of advertising occurs through television, a medium often criticized for its lack of informational content and intent to persuade the consumer. This supports hypotheses offered by the signalling literature that it is the amount as opposed to the content of advertising that is important in providing information, especially for experience goods. Television provides an efficient medium to reveal that large advertising outlays are being made and consequently that quality is high.(19)

The above study is by no means a definitive test of the signalling hypothesis. While it is advantageous to examine a single industry, other industries should be examined to determine the generality of the conclusions offered here. Moreover, the present study only focuses on advertising's ability to signal quality. Examining the role of price and/or warranties would also be beneficial.

I have benefited from discussions with Kathleen Carroll, Gary Fournier, Jill Hendrickson, Brad Kamp, Leola Ross, and Tim Sass. I would also like to thank two anonymous referees and Editor Jonathan Hamilton for their helpful suggestions. Any remaining errors are solely my responsibility.

1 See, however, Hertzendorf (1990) and Overgaard (1991), who allow for advertising that signals quality and has a direct impact on demand.

2 Data on cable television advertising were reported starting in 1985, while figures for newspapers were not available until 1987.

3 CR rates vehicles for five years beginning in the year following its introduction. [q.sub.j,t] is the average of these five trouble indices. For example. for the 1985 Ford Thunderbird, trouble indices were available beginning in 1986 and reported through 1990. Other measures were tested but had no appreciable impact on the results reported below. These included using the first available trouble index, the most favorable rating, and the least favorable rating.

4 The 1985 Ford Thunderbird is rated average (a value of three) in all five years between 1986 and 1990. In contrast, the 1985 Buick Regal has ratings of 5, 5, 3, 4, 4 between 1986 and 1990. respectively. The average trouble index, the average of these five ratings, for the Ford Thunderbird is 3, whereas for the Buick Regal it is 4.2. The higher value of the average trouble index for the 1985 Regal classifies it as a lower quality vehicle than the 1985 Thunderbird.

5 Of course, advertising also influences sales, so the use of sales as an explanatory variable introduces the potential for simultaneity bias. This is addressed using two-stage least squares to estimate the results reported in section 4.

6 Average transaction price for model j at time t is not available. The price at time t + 1, that is, the price of the model when it is one year old. rather than the manufacturer's suggested retail price (MSRP), is used for two reasons: First, it is likely to be highly correlated with the new car price but avoids the potential simultaneity bias resulting from using the price at time t. Second. it will reflect the market's perceptions about reliability, popularity, etc., that the time t transaction price captures but that the MSRP may not. Using the MSRP did not, however, significantly affect the estimated results reported below.

7 The structural equations for [SALES.sub.j,t] and [ADV.sub.j,t] forming the basis for the two-stage least squares estimation are as follows:

[SALES.sub.j,t] = [[Beta].sub.0] + [[Beta].sub.1][ADV.sub.j,t] + [[Beta].sub.2][SALES.sub.j,t-1] + [[Beta].sub.3][AVERAGE TROUBLE INDEX.sub.j,t] + [[Beta].sub.4][STYLE CHANGE.sub.j,t] + [[Beta].sub.5][NEW ENGINEERING.sub.j,t] + [[Beta].sub.6][INTRODUCTORY MODEL.sub.j,t] + [[Beta].sub.7][RIVAL STYLE CHANGE.sub.j,t] + [[Beta].sub.8][RIVAL ENGINEERING CHANGE.sub.j,t] + [[Beta].sub.9]TREND + [[Beta].sub.10]GM + [[Beta].sub.11]FORD + [[Beta].sub.12]CHRYSLER + [[Beta].sub.13]COMPACT + [[Beta].sub.14]MEDIUM + [[Beta].sub.15]LARGE + [[Beta].sub.16][GAS PRICE.sub.t] + [[Beta].sub.17][REAL PER CAPITA INCOME.sub.t] + [[Epsilon].sub.j,t]

[ADV.sub.j,t] = [[Alpha].sub.0] + [[Alpha].sub.1][SALES.sub.j,t] + [[Alpha].sub.2][ADV.sub.j,t-1] + [[Alpha].sub.3][AVERAGE TROUBLE INDEX.sub.j,t] + [[Alpha].sub.4][STYLE CHANGE.sub.j,t] + [[Alpha].sub.5][NEW ENGINEERING.sub.j,t] + [[Alpha].sub.6][INTRODUCTORY MODEL.sub.j,t] + [[Alpha].sub.7][RIVAL STYLE CHANGE.sub.j,t] + [[Alpha].sub.8][RIVAL ENGINEERING CHANGE.sub.j,t] + [[Alpha].sub.9]TREND + [[Alpha].sub.10]GM + [[Alpha].sub.11]FORD + [[Alpha].sub.12]CHRYSLER + [[Alpha].sub.13]COMPACT + [[Alpha].sub.14]MEDIUM + [[Alpha].sub.15]LARGE + [[Alpha].sub.16] [NUMBER OF MODELS.sub.j,t] + [[Alpha].sub.17][PRICE.sub.j,t] + [[Upsilon].sub.j,t]

8 The high partial [R.sup.2] and F-statistic on the excluded instruments ([SALES.sub.j,t-1], [GASPRICE.sub.t], and [REALPCY.sub.t]) suggest a strong correlation between the instruments and endogenous explanatory variable ([SALES.sub.j,t],) and negligible finite-sample bias. The partial [R.sup.2] is that from regressing [SALES.sub.j,t], on the excluded instruments, while the F is found by restricting the coefficients on the excluded instruments to zero in the first-stage regression (see Bound, Jaeger, and Baker [1995]). Furthermore, a Hausman test on the validity of the excluded instruments (comparing the OLS estimate of the coefficient on [SALES.sub.j,t] with its IV estimate, that is. testing if plim([[Beta].sup.IV] - [[Beta].sup.OLS]) = 0) yields a test statistic of [Mathematical Expression Omitted], failing to reject the null of no regressor-error correlation.

9 The Box-Cox transformation may be applied to some or all of the variables in the equation and allows the functional form to be determined by the data. Applying the transformation to all variables involves maximum likelihood estimation of the following: [Mathematical Expression Omitted], where [Y.sup.[Lambda]] = ([Y.sup.[Lambda]] - 1) and [Mathematical Expression Omitted] for all i. [Lambda] = 1 results in the linear model, whereas [Lambda] = 0 results in the log-linear model. By allowing [Lambda] to vary for each variable, more flexible specifications arise.

10 The specification in Equation 2 (the lagged effects model) was tested and favored over a model where lagged advertising is excluded but the errors are serially correlated (the current effects model). Distinguishing between these two involves running the following regression:

[Mathematical Expression Omitted],

where Z includes [q.sub.j,t], [K.sub.j,t], [C.sub.j,t], and [X.sub.j,t], [Rho] is the autocorrelation coefficient, and [[Mu].sub.j,t] = [Rho][[Mu].sub.j,t-1] + [[Epsilon].sub.j,t]. The lagged effects model is supported by a statistically insignificant coefficient on lagged sales (t-statistic was 1.10). See Griliches (1967) for more detail.

11 For an excellent summary of improving quality between 1980 and 1990, see "Is Detroit Closing the Reliability Gap?" Consumer Reports, April 1991, pp. 248-249.

12 Measuring expenditures as advertising/sales or total advertising by class of car yields a similar pattern, with the exception that U.S. producers have greater expenditures than foreign producers in the large and medium class.

13 Resnick and Stern (1977) show that only 49% of television advertisements were informative, defined as communicating 1 of 14 informational cues (e.g., price, performance, availability). When required to communicate 2 and 3 of the 14 information cues, the percent of television advertisements deemed informative fell to 16% and 1%, respectively.

14 While the coefficient on quality is statistically significant, most of the model's explanatory power is captured by lagged advertising. Dropping quality from the estimated equation lowers the [R.sup.2] from 0.66 to 0.65. Estimating the model without lagged advertising, however, yielded similar results, with the [R.sup.2] dropping from 0.30 to 0.29 when quality is excluded.

15 An F-test that all manufacturer and class-size dummies are jointly equal to zero resulted in an F-statistic of 2.97, rejecting the null that all are jointly equal to zero. Excluding GM, however, resulted in an F of 1.40, failing to reject the null.

16 For introductory models where no previous models exist, the manufacturer's average trouble index was used.

17 Equation 2 was also estimated with the average trouble index from the previous model replacing the (future) average trouble index. The coefficient was negative but insignificant, reaffirming that producers are not simply advertising to tout high quality in previous years and lending further credence to the signalling hypothesis.

18 Again, since no earlier data exist for introductory models, their improvement in quality is calculated as the change from the manufacturer's average trouble index in the previous year.

19 While advertising prices vary by time of day, show, network, etc., on average, a 30-second commercial on network television in 1996 cost $169,000 (Advertising Age, September 16, 1996). Expressing this in real dollars, assuming that the average expenditures given in Table 3 were on an average quality vehicle, and using the 19% difference between quality classes given in Table 4 implies 40.37, 35.23, 41.79, 60.34, 63.34, and 71.09 more commercials per year for a vehicle of much better than average quality (CR rank = 1) and a much worse than average quality vehicle (CR rank = 5) for the years 1985-1990, respectively.

References

Advertising Age. 1996. September 16.

Bagwell, Kyle, and Garey Ramey. 1993. Advertising as information: Matching products to buyers. Journal of Economics and Management Strategy 2:199-243.

Berry, Steven, James Levinsohn, and Ariel Pakes. 1995. Automobile prices in market equilibrium. Econometrica 63: 841-90.

Bound, John, David A. Jaeger, and Regina M. Baker. 1995. Problems with instrumental variables estimation when the correlation between the instruments and the endogenous explanatory variable is weak. Journal of the American Statistical Association 90:443-50.

Box, G. E. P., and D. R. Cox. 1964. An analysis of transformations. Journal of the Royal Statistical Society, (Series B) 26:211-43.

Caves, Richard E. 1986. Information structures of product markets. Economic Inquiry 24:195-212.

Caves, Richard E., and David P. Greene. 1996. Brands' quality levels, prices, and advertising outlays: Empirical evidence on signals and information costs. International Journal of Industrial Organization 14:29-52.

Consumer's Union. 1986-1995. Consumer Reports. New York: Consumer's Union.

Crain Communications, Inc. 1996. Advertising Age, 12 February. Chicago: Crain Communications.

Crocker, Keith J. 1986. A reexamination of the 'lemons' market when warranties are not prepurchase quality signals. Information Economics and Policy 2:147-61.

Dorfman, Robert, and Peter O. Steiner. 1954. Optimal advertising and optimal quality. American Economic Review 44: 826-36.

Ehrlich, Isaac, and Lawrence Fisher. 1982. The derived demand for advertising: A theoretical and empirical investigation. American Economic Review 72:366-88.

Griliches, Zvi. 1967. Distributed lags: A survey. Econometrica 35:16-49.

Hertzendorf, Mark N, 1990. I'm not a high-quality firm - But I play one on TV: A model of signaling product quality. Unpublished paper, University of Rochester.

Hertzendorf, Mark N. 1993. I'm not a high-quality firm - But I play one on TV. Rand Journal of Economics 24:236-47.

Khilstrom, Richard E., and Michael H. Riordan. 1984. Advertising as a signal. Journal of Political Economy 92:427-50.

Klein, Benjamin, and Keith B. Leffler. 1981. The role of market forces in assuring contractual performance. Journal of Political Economy 89:615-41.

Kwoka, John E. Jr. 1993. The sales and competitive effects of styling and advertising practices in the U.S. auto industry. The Review of Economics and Statistics 75:649-56.

Leading National Advertisers, Inc. 1984-1990. Class/brand tear to date. New York: Leading National Advertisers, Inc.

Leffler, Keith B. 1981. Persuasion or information? The economics of prescription drug advertising. Journal of Law and Economics 24:45-74.

Lutz, Nancy A. 1989. Warranties as signals under consumer moral hazard. Rand Journal of Economics 20:239-55.

Milgrom, Paul, and John Roberts. 1986. Price and advertising signals of product quality. Journal of Political Economy 94:796-821.

Nelson, Phillip. 1970. Information and consumer behavior, Journal of Political Economy 78:311-29.

Nelson, Phillip. 1974. Advertising as information. Journal of Political Economy 82:729-54.

Overgaard, Per Baltzer. 1991. On the nature of advertising and its role as a signal of quality. Unpublished paper, University of Aarhus.

Porter, Michael E. 1976. Interbrand choice, media mix, and market performance. American Economic Review: Papers and Proceedings 66:398-406.

Resnick, Alan, and Bruce L. Stern. 1977. An analysis of information content in television advertising. Journal of Marketing 41:50-3.

Schmalensee, Richard. 1978. A model of advertising and product quality. Journal of Political Economy 86:485-503.

Tellis, Gerard J., and Claes Fornell. 1988. The relationship between advertising and product quality over the product life cycle: A contingency theory. Journal of Marketing Research 25:64-71.