An empirical examination of quality certification in a "lemons market".

Wimmer, Bradley S. ; Chezum, Brian

I. INTRODUCTION

Researchers have long understood that a variety of mechanisms may be used to overcome problems of adverse selection. Stigler (1961, 224) notes that "[s]ome forms of economic organization may be explicable chiefly as devices for eliminating uncertainties in quality," and Akerlof (1970) cites independent groups, such as the Consumers Union and United Laboratories, that test and certify the quality of goods. There is, however, little empirical evidence that illustrates the effectiveness of such mechanisms. In this article we examine the effect certification has on a market characterized by adverse selection. Using data from the market for young thoroughbreds, we compare the performance of sales where auction houses provide certification services with sales where certification is absent.

The thoroughbred racehorse market consists of two distinct types of public auctions: certified and noncertified sales. In a certified sale, auction houses physically inspect the horses nominated to their sales, selling only the horses they conclude are from the upper end of the quality distribution. In noncertified sales, auction houses sell all horses nominated to a sale. Our empirical strategy is to perform tests that indicate whether adverse selection affects market outcomes in either certified or noncertified sales. A finding that adverse selection is present in noncertified sales but absent in certified sales is consistent with the hypothesis that certification alleviates problems of adverse selection.

We adopt three approaches to test for the presence of adverse selection. (1) The first approach is in the spirit of Chiappori and Salanie (2000), who examine the relationship between unobservable factors from participation and performance equations. We model adverse selection as a case of sample-selection bias and examine the correlation between errors in participation and price equations. (2) To accomplish this we use a unique data set that allows us to estimate how breeder decisions to sell or retain horses affect market prices. The data set consists of a 10% random sample of all thoroughbreds born in 1993 and includes both horses retained by their breeders, who own a horse at the time of its birth, and horses that breeders chose to sell.

Our second test follows Genesove (1993) and Chezum and Wimmer (1997) by examining the relationship between seller characteristics and price. Theoretically, when observable seller characteristics are correlated with seller incentives to select goods adversely, prices should reflect these differences. Chezum and Wimmer observe that some breeders sell all of their horses, whereas others retain a portion to race. They find that market prices are inversely related to the extent of a breeder's involvement in racing, concluding that adverse selection affects market outcomes. We extend this work in two respects. First, we isolate the effect seller characteristics have on price through the decision to sell or retain a horse. Second, we compare the effect seller characteristics have on observed prices in certified and noncertified sales.

Our last test follows Bond (1982) who attempts to identify adverse selection in the market for used trucks by comparing the repair records of trucks that were sold with the records of trucks retained by their original owners. Because the horses in our sample had not begun their racing careers at the time they were sold, we use data on racetrack earnings as an ex-post measure of quality and compare the quality of horses sold in noncertified sales with certified horses and horses retained by their breeders.

The results from each approach are consistent with certification alleviating problems of adverse selection. We find that holding observable attributes constant, horses that would receive unusually high market prices in noncertified sales are even more valuable in other options and are less likely to be sold in a noncertified sale. In certified sales, we find no evidence of adverse selection. We show that the difference in expected prices between certified and noncertified sales can be attributed primarily to the selection process. Finally, we find that holding other factors constant, horses sold in noncertified sales earned less money from racing than horses retained by their breeders. No significant difference in earnings is found between retained and certified horses. Overall, our results indicate that adverse selection affects market outcomes in noncertified sales but plays no significant role in certified sales.

The remainder of the article is organized in the following manner. Section II summarizes the certification process for young thoroughbreds and reviews the literature on certification. Section III presents the empirical strategy. Section IV discusses the data. Results are presented in section V, and section VI concludes.

II. ADVERSE SELECTION, CERTIFICATION, AND THE THOROUGHBRED INDUSTRY

When sellers have an informational advantage over buyers, trades that would be mutually beneficial under full information may not be consummated. This market failure creates a profit opportunity for certifying agents who can provide buyers with credible information about the quality of goods sold. Biglaiser (1993) shows that certifying agents, or middlemen, receive a larger return on investments in expertise than do occasional buyers because middlemen inspect a relatively larger number of goods. He also shows that the value created by middlemen is increasing in the proportion of goods that are of a relatively low quality and the spread in the value between high and low-quality goods. (3) Certification will be used when it provides buyers with information that is either unavailable or more costly to obtain when certification is not available.

Certification services are likely to be valuable in the market for thoroughbred racing prospects because breeders are likely to have an informational advantage over potential buyers and the spread between the value of low- and high-quality thoroughbreds is large. Chezum and Wimmer (1997) argue that breeders, having raised their thoroughbreds since birth, have access to information on a horse's full medical history, temperament, physical attributes, and other information that is not readily available to the general public. Such information is likely to be valuable because the difference in value between a low- and high-quality horse is large. Less than 1% of all thoroughbreds win the highest-quality stakes races, and many never earn a dollar from racing. (4)

Auction houses are well positioned to provide certification services because the majority of young thoroughbreds are sold by a limited number of auction houses. (5) For their services, auction houses receive a commission that is generally equal to a nonrefundable fee (typically $500 or $1,000) or 5% of the sales price, whichever is greater. Auction houses have the incentive to alleviate problems of adverse selection because their revenues are directly related to the number and quality of horses sold. It is not surprising that auction houses certify the quality of a portion of the horses they sell.

The certification process begins with thoroughbred owners filing a nomination form and paying the nonrefundable fee to the auction house. (6) The nomination form includes information on the horse's pedigree and, in many cases, the breeder's estimate of the price a horse will fetch at auction. The auction house ranks nominated horses by estimated prices and quality of pedigrees, eliminating horses with a relatively low rank. (7) The remaining horses are physically inspected at the breeder's farm by a panel of experts. (8) The panel studies the physical appearance of the horses, examining them for soundness and other characteristics that indicate potential success as a racehorse. Horses deemed to be of a relatively high quality are sold in certified sales. The auction house does not reveal the breeder's initial price estimate or any of the specifics of its inspection.

To alleviate adverse selection, certification must provide buyers with information that is not readily available elsewhere. Thoroughbred certification provides buyers with information about the quality of horses through both a direct and an indirect channel. The direct channel comes from the auction house's physical examination of horses, which is consistent with services performed by expert middlemen. The indirect channel arises because auction houses use information provided by breeders when certifying horses. Because auction houses have the incentive and ability to punish breeders who misrepresent the quality of their horses, auction houses provide breeders with an avenue to reveal private information credibly.

Biglaiser and Friedman (1994) argue that multiproduct intermediaries suffer a "reputational spillover" if they continue selling goods from producers that misrepresent quality. When a multiproduct intermediary receives consumer complaints about the quality of a producer's goods it has the incentive to pull this producer's goods from its shelves. (9) The auction house's requirement that breeders provide it with an estimate of a horse's value is consistent with this function. Auction houses sell horses from many different breeders and have the incentive to protect their reputation by punishing breeders who misrepresent quality. Because information on the quality of horses is revealed through the inspection process and during the horse's racing career, auction houses can monitor the accuracy of breeder reports and exclude from certified sales breeders who develop a reputation for misrepresenting quality. This indirect mechanism induces breeders to provide the auction house with accurate information about the qual ity of their horses, reducing asymmetries of information.

Because auction houses provide information only on whether or not a horse is certified, certification is similar to Leland's (1979) minimum quality standard. (10) A minimum quality standard truncates the distribution of quality from below, raises prices, and induces seller to offer higher quality horses, but adverse selection may still exist in certified sales. Because buyers base price offers on the expected quality of horses sold in a lemons market, the breeder's optimal strategy is to sell relatively low-quality horses and retain the remainder. This may not be the case if buyers inspect horses more intensively in certified sales.

Potential buyers are allowed to inspect horses before all thoroughbred sales and may perform minimally intrusive tests. (11) Barzel (1982) argues that sellers have the incentive to sort products to reduce the amount of resources devoted to information acquisition. If the expected quality of horses is increasing in observable attributes (such as pedigree) inspection expenditures are likely to be increasing in these attributes. By eliminating inferior specimens from certified sales, auction houses allow buyers to target their inspection efforts on certified horses. In addition, the auction house's initial sorting and subsequent inspection by potential buyers serves to dissuade breeders from nominating low-quality horses to certified sales. Although we are unable to observe buyer inspection expenditures, certification reduces these expenditures by publicly revealing the results of their initial inspection, thereby alleviating problems of asymmetric information.

III. EMPIRICAL STRATEGY

The market for young thoroughbreds provides a natural test of the effect certification has on market outcomes. We test for the presence of adverse selection in both segments of the market, namely, certified and noncertified sales. A finding that adverse selection affects outcomes in noncertified sales but has no effect in certified sales is consistent with the hypothesis that certification alleviates problems of adverse selection.

To identify adverse selection we observe that when breeders adversely select the horses they sell, the sample of horses sold is censored systematically, which, as shown by Heckman (1979), results in sample-selection bias. A test for adverse selection is that selectivity bias causes observed prices to be lower than the prices generated from a random selection process. (12) This approach explicitly estimates the effect breeder-selection decisions have on observed prices.

Thoroughbred breeders have the option to sell their horses at one of several auctions. (13) To keep the empirical analysis tractable, we categorize sales as either certified or noncertified and assume that breeders choose among the following three options: retain a horse, R; sell a horse in a noncertified sale, NC; or nominate a horse to a certified sale, C. Because the model consists of a trichotomous choice and two price equations, we use Lee's (1983) methodology to estimate a model that allows for selectivity bias in a polychotomous-choice setting.

Breeders are assumed to choose the option that maximizes utility. (14) The maximum utility attainable from each option is treated as a function of market prices in certified and noncertified sales, a horse's expected earnings net of costs, (15) the cost of taking a horse to sale and breeder characteristics. As discussed by Chezum and Wimmer (1997) some breeders raise horses for the express purpose of selling them, and others retain a portion to race. To capture the effect differences in breeder characteristics have on indirect utility, the variable Racing Intensity, RI, is included. RI captures the extent of a breeder's racing operation. We assume that buyers observe each breeder's RI and that the utility a breeder receives from retaining a horse is increasing in RI.

The maximum utility breeder i receives from option j (j = R, NC, C) is given by [V.sub.ji] = V ([P.sub.ji], [K.sub.ji], [RI.sub.i]) where V (.) is the indirect utility function, [P.sub.j] is the price in a certified or noncertified sale or net earnings if retained, (16) [K.sub.j] is the monetary cost of option j, and [RI.sub.i] captures differences in breeders.

The natural log of [P.sub.j] is assumed to be a linear function of a thoroughbred's observable characteristics:

(1) [P.sub.ji] = [[beta]'.sub.j][X.sub.i] + [e.sub.ji],

where [e.sub.ji] is the random component for observation i in option j and [[beta]'.sub.j][X.sub.i] as a vector of a horse's observable attributes and the associated coefficients.

Following Trost and Lee (1984) we assume that the indirect utility function is linear and can be decomposed into a nonstochastic and random component:

(2) [V.sub.ji] = [[gamma]'.sub.j][Z.sub.i] + [u.sub.ji],

where [u.sub.ji] is the random component for observation i in option J, [Z.sub.i] is a vector of observable variables, and [[gamma].sub.j] is the associated vector of coefficients. Although the indirect utility a breeder receives from each option is not observed, final outcomes of the selection process are. Prices are observed when selling provides the highest indirect utility; [V.sub.si], > max [V.sub.ji], (s = C, NC; j = R, NC, C, [not equal to] s).

To estimate the probability that an option is chosen we assume that [u.sub.ji] follows an extreme value distribution and estimate a multinomial logit choice model. The probability that breeder i chooses option s is given by: (17)

(3) [P.sub.s] ([[gamma]'.sub.j][Z.sub.i] = exp ([[gamma]'.sub.s][Z.sub.i] / [1 + [[sigma].sub.j=NC, c] exp([[gamma]'.sub.j][Z.sub.i])],

where the retained option is used as the base category.

If horses are adversely selected into a sale, unobservable factors that increase prices reduce the probability that a horse is sold, resulting in selectivity bias in the price equation. Heckman (1979) shows that this is a classic case of omitted variable bias and that consistent estimates can be obtained by including the inverse Mills ratio to account for the incidental truncation caused by the selection process.

As shown by Lee (1983), the inverse Mills ratio for a polychotomous-choice model is calculated by transforming the estimated probabilities in the following manner:

(4) [[lambda].sub.s] ([[gamma]'.sub.j]Z) = [phi][[[PHI].sup.-1] ([P.sub.s]([[gamma]'.sub.j][Z.sub.i])]/[P.sub.s]([[gamma]'.sub.j][Z.s ub.i]),

where [phi] is the standard normal density function and [PHI] is the standard normal distribution function. (18) We include the appropriate inverse Mills ratio in the certified and noncertified price equations to control for selection bias to obtain unbiased estimates of [[beta].sub.c] and [[beta].sub.NC]:

(5) [P.sub.ci] = [[beta]'.sub.C][X.sub.i] + [[beta].sub.[lambda]C][[lambda].sub.C]([[gamma]'.sub.j]Z) + [[epsilon].sub.Ci]

[P.sub.NCi] = [[beta]'.sub.NC][X.sub.i] + [[beta].sub.[lambda]NC][[lambda].sub.NC]([[gamma]'.sub.j]Z) + [[epsilon].sub.NCi]

where [[epsilon].sub.Ci] and [[epsilon].sub.NCi] are the random disturbance terms. (19)

Our first test for adverse selection focuses on the sign of the estimated coefficient on inverse Mills ratios contained in equation (5). Lee (1983) shows that the sign of coefficient on the inverse Mills ratio represents the covariance of the errors from the price and selection equations. (20) A negative coefficient indicates that horses that would receive unusually high market prices, given their observed attributes, are even more valuable in their next-best options and are less likely to be sold in that sale category. Such a finding is consistent with adverse selection.2' By contrast, a positive coefficient on the inverse Mills ratio is indicative of positive selection, whereas a coefficient that is not statistically different from zero indicates that selection bias does not affect the sample.

The second test incorporates Genesove's (1993) test for adverse selection into the selection-bias framework. Genesove hypothesizes that when seller incentives to select goods adversely are correlated with observable seller characteristics, prices will reflect these differences. To motivate trade when information is asymmetric, Genesove argues that sellers are capacity constrained and that it is optimal for sellers to sell low-quality goods to relieve capacity constraints. Using RI to proxy the severity of capacity constraints, Chezum and Wimmer (1997) apply Genesove's approach to the market thoroughbred yearlings and find that prices are inversely related to racing intensity. (22)

In the framework developed, an increase in RI increases the indirect utility a breeder receives from retaining a horse, thereby decreasing the probability that a horse is sold. If breeders adversely select the horses they sell, a reduction in the probability that a horse is sold results in lower prices. To measure the effect RI has on price, through the breeder's selection decision, we calculate [[beta].sub.[lambda]s]([partial][[lambda].sub.s]/[partial]RI), which is the effect RI has on prices through the inverse Mills ratio. (23)

This second approach suggests two tests for the effect certification has on market performance. First, we expect to find that an increase in RI reduces prices through the mechanism already defined when goods are adversely selected. A finding that an increase in RI reduces prices in noncertified sales but has no effect in certified sales is consistent with certification alleviating problems of adverse selection. Second, we examine the effect RI has on the probability that a horse is sold. In general, we expect to find an inverse relation between RI and the probability that a horse is sold. A finding that RI has no effect on the probability that a horse is sold in a certified sale indicates that certification affects breeder decisions to sell or retain a horse and is evidence that certification attracts horses to sales that would have been retained if certification were not available. (24)

To estimate the magnitude of the effect adverse selection has on price we follow Lee (1995) who shows how estimates from the corrected price regressions can be used to estimate the opportunity cost of a selection decision. Lee shows that when using a logit model for selection equations the opportunity cost of a choice is based only on the probability that an option is not chosen. Following Lee, we calculate the expected price generated by a horse randomly allocated to sale s, which is defined as E[P.sub.s]\X, Z], s = C, NC. We also calculate the expected price of a horse, given it was selected into sale s', E[[P.sub.s]\X, Z, s = s'], and the expected price a horse that was not offered in sale s' would receive if sold in sale s', E[P.sub.s]\X, Z, s [not equal to] s']. (25) To estimate the magnitude of adverse selection, we calculate the difference between the expected price of a random horse in a noncertified sale and the expected price of a horse selected into a noncertified sale, E[P.sub.s]\X, Z] - E[P.sub.s ]\X, Z, s = NC].

Our third test uses racetrack earnings per race as a proxy for quality. (26) This approach follows Bond (1982) by using an ex-post measure of quality to identify adverse selection. A finding that noncertified horses earn less than horses retained or certified, holding other factors constant, is consistent with the presence of adverse selection in noncertified sales and provides additional evidence that certification alleviates problems of adverse selection.

IV. DATA

This study uses an original data set that consists of a 10% random sample of all thoroughbreds born in the United States in 1993. The data include thoroughbreds retained by their original owners as well as thoroughbreds taken to auction. These data were obtained from The Jockey Club's Foals of 1993 (1995) (Foal Book), an annual supplement to the American Stud Book, which, includes information on all thoroughbreds born and registered in the United States, Canada, and Puerto Rico.(27) The sample consists of all U.S.-born horses listed on every tenth page of the Foal Book and includes 3,376 horses. (28)

Sales data for each observation were obtained by matching data from the Foal Book to information contained in the Blood Horse's Auction Guide. The Auction Guide includes the results of every public thoroughbred auction held in North America. These data allow us to identify the horses that were offered for sale at public auction and the prices they received. The empirical analysis follows Chezum and Wimmer (1997) and includes horses offered for sale as yearlings or younger. (29) In the sample, 925 horses were offered for sale as yearlings or younger at least once. The empirical analysis focuses on the initial sale, although horses may be sold more than once. In the data 159 of the 925 horses sold (17.2%) were marketed in certified sales. The remaining horses were sold in noncertified sales.

To distinguish sellers, we use the variable Racing Intensity. For each racing season the American Racing Manual (ARM) (1993, 1994) publishes the earnings of thoroughbred owners whose horses earned at least $50,000 and whose horses they bred earned at least $30,000 in that year. To measure Racing Intensity 1993 data were gathered for each breeder on the number of races started by horses they owned at the time of a race, Racing Starts, and the number of races that horses they bred started, Breeder Starts. (30) Racing Intensity is equal to the ratio of Racing Starts to Breeding Starts + 1. The lower limits for inclusion in the ARM result in relatively small breeding operations not appearing in the ARM data. To account for this we include a variable, Unlisted Breeder, that is set equal to one if both Racing Starts and Breeding Starts are unobserved and zero otherwise.

Data on a variety of covariates are obtained from the ARM and from Blood-stock Research Information Service's American Produce Records (2000). We gather information to control for the age and gender of the thoroughbred, the number siblings, the success of the sire (father) and dam (mother) both at the track and in the breeding shed, and finally information to control for the costs associated with going to auction.

For each observation we include the Age in Months, which is the age of the horse, measured in months, on 1 January 1994 (the horses in the sample were born in 1993) in the participation equation. For the price equation we include we include Age at Sale, which is also measured in months. Older horses should be more likely to be sold and will receive higher prices at sale. Colt is set equal to one for male thoroughbreds and zero for female thoroughbreds.

Finnochio (1995) shows that expected quality of a horse is a decreasing in the number of foals previously produced by the thoroughbred's dam (mother). Siblings is defined as the number of horses previously produced by a horse's dam at the time of an observation's birth.

To account for the quality of a thoroughbred's pedigree variables on the quality of the sire and dam both as racehorses and in breeding are included. Dam Stakes Winner and Sire Grade I Winner are set equal to one if the dam (sire) won a stakes (Grade I stakes) race and zero otherwise.

The quality of the dam and sire in breeding is proxied by variables that capture their offspring's quality. For the dam we include Stakes-Winning Siblings, which equals the number of a dam's offspring that have won a stakes race. For the sire we calculate the natural log of a sire's offspring's earnings per performer in 1992, Log Earnings per Sire Offspring. The ARM does not include information on sires whose offspring earned less than $50,000 in 1993. The variable Unlisted Sire, which is set equal to one if the sire is not in the ARM data and zero otherwise, is included in the regressions. (31) Finally, the number of foals produced by a sire should be increasing in the demand for a sire's services. We include the variable Sire's Crop, which is equal to the number of foals produced by a sire in 1993.

To identify the selection equation a set of state-bred dummy variables that proxy the cost of taking a horse to a sale is used. In our data certified sales are conducted in Kentucky, Florida, New York, and California. The qualitative variable Certified State is defined as one if the horse was born in any of these states and zero otherwise. Similarly, Sale State is set equal to one if there is a noncertified sale in the state where the horse was born and zero otherwise.

Table 1 contains the summary statistics for the variables used in the analysis. The first column contains descriptive statistics for the entire sample, and the second reports summary statistics for horses not sold as yearlings. The final two columns report descriptive statistics for horses sold in noncertified and certified sales, respectively. In the sample, approximately 27% of the horses in the sample were sold as yearlings. The average price in a noncertified sale is $13,268, whereas certified horses fetched an average price of price of $93,437, a difference of over 600%.

In terms of earnings, horses sold in both certified and noncertified sales on average earn more than horses that were not sold. (32) The average certified horse earned $48,475, with noncertified horses earning an average of $27,188, a difference of 78.3%. Horses that are not sold earned an average of $17,716. Many of the horses in our sample never started a race. Over 80% of the certified and noncertified horses started a race, but only 69% of the horses not sold started a race. It is likely that some horses in our data were bred for reasons other than racing. Also, clearly inferior horses may be easy to identify and not sold. Our data do not allow us to identify such horses. (33)

The mean Racing Intensity of retained horses is larger than the averages for horses sold in certified and noncertified sales, and Racing Intensity's mean is slightly higher in certified than in noncertified sales. The average quality of a horse's sire and dam is relatively higher in certified sales than noncertified sales, as shown by the larger means for Sire Grade I Winner, Dam Stakes Winner, Stakes-Winning Siblings, Sire's Crop, and Log Earnings per Sire Offspring. The means of these variables for retained horses are lower than found in noncertified sales. We also see that a higher percentage of horses not sold are from sires not included in the ARM and from relatively small breeders. Some of these breeders may raise horses for reasons other than racing, which is consistent with horses not sold being less likely to reach the races.

In the full sample, 47% of the horses were born in states where a certified sale is held, whereas 89% of the horses sold in certified sales were born in these states. For noncertifled sales, 82% of the full sample were born in a state with a sale, and 91% of the horses sold were born in these states.

V. EMPIRICAL RESULTS

Table 2 contains results from the empirical analysis. The first two columns display the results from the price regressions as shown in equation (5). Columns three and four report the results from the multinomial logit model. Estimates of the marginal effects and their associated z statistics evaluated at the sample mean are reported. The last column provides the results from the analysis of U.S. earnings-per-race data.

The primary focus in the price regressions is the effect selection has on observed prices. The coefficient on the Inverse Mills Ratio is negative and significant in the noncertified regression. This finding indicates that, holding observable attributes constant, unobservable factors that increase the probability of a horse being sold in a noncertified sale are inversely related to unobservable factors that increase price. In certified sales the coefficient on the Inverse Mills Ratio is negative and highly insignificant. (34) The results suggest adverse selection is present in noncertified sales but has no effect on prices in certified sales and are consistent with certification alleviating problems of adverse selection.

The results also indicate that horses sold by racing-intensive breeders receive lower prices in noncertified sales. The sample estimate of [[beta].sub.[lambda]s]([partial][[lambda].sub.s]/[partial]RI), evaluated at the sample mean, is negative and marginally significant in the noncertified sale regression. (35) In certified sales, the estimate of [[beta].sub.[lambda]s]([partial][[lambda].sub.s]/[partial]RI) is insignificant. These results are generally consistent with Chezum and Wimmer's (1997) earlier finding that racing-intensive breeders are more likely to adversely select the horses they sell but show that differences in seller characteristics have no effect on prices in certified sales.

The results for the remaining variables are as expected. In general, observable characteristics are directly related to prices. The coefficients on the majority of these variables have the expected signs and are statistically significant. (36)

The results from the multinomial logit show that racing-intensive breeders are less likely to offer their horses for sale in noncertified sales. An increase in Racing Intensity results in a statistically significant decrease in the probability that a horse is sold in a noncertified sale. In certified sales, the marginal effect of Racing intensity is not statistically different from zero. (37) These findings are consistent with the presence of adverse selection in noncertified sales and that certification alleviates problems of adverse selection.

The results from the multinomial logit indicate that observable characteristics play an important role in both certified and noncertified sales. The marginal effects of Sire Grade I Winner and Sire's Crop are positive and statistically significant in both the certified and noncertified regressions. Log Earnings per Sire Offspring has a positive and statistically significant effect in the certified sale equation, but is insignificant in the noncertified sale equation. Only the marginal effect of Stakes-Winning Sibling increases the probability of certification but decreases the probability that a horse is sold in a noncertified sale. The marginal effect of Stakes-Winning Sibling is statistically significant and negative in the noncertified equation and is positive (but of only marginal significance) in the certified equation. (38) These results provide some insight into the certification process. Observable attributes appear to affect the probability of inclusion in both sales, suggesting that certifying agent s use more than the quality of a horse's pedigree when deciding to certify a horse.

To estimate the effect adverse selection has on prices in noncertified sales, we estimate the expected price of a horse randomly allocated to a noncertified sale, E[[P.sub.NC]\X, Z], and the expected price of a horse given it was sold in a noncertified sale, E[[P.sub.NC]\X, Z, NC = 1], using Lee's (1995) methodology. For the entire sample, we calculate that the average predicted value of E[[P.sub.NC]\X, Z] is $88,259, whereas the average for E[[P.sub.NC]\X, Z, NC = 1] is $9,253, nearly a tenfold difference. (39) Although the magnitude is somewhat surprising, it shows that adverse selection not only affects market outcomes but also that its effect is quite severe.

To estimate the effect certification has on prices, which may include differences in factors observed by the market but not the econometrician, we estimate E[[P.sub.C]\X, Z] - E[[P.sub.NC]\X, Z, NC = 1], where the expected price of a random horse in a certified sale is used because self-selection does not affect prices in these sales. We find that E[[P.sub.C ]X, Z] - E[[P.sub.NC]\X, Z, NC = 1] = $99,122 - $9,253 = $89,869. Approximately 13% of this difference can be attributed to differences in the value buyer's place on observable attributes, and the remaining 87% is attributable to the selection process. (40)

The final column reports the results for the earnings-per-race equation. (41) The results show that noncertified horses earn less money at the track than horses that were not sold as yearlings. The coefficient on Non-Certified is negative and statistically significant. The coefficient on Certified is also negative but does not reach statistical significance. These results reinforce the earlier findings that adverse selection affects the market for noncertified young thoroughbreds.

VI. CONCLUSIONS

In this article we show that certification alleviates problems of adverse selection in thoroughbred auctions by examining the effect certification has on breeder decisions to retain or sell horses and the effect these decisions have on prices. In addition, we compare the performance of horses sold in certified and noncertified sales with horses retained by their breeders. In each case the evidence indicates that certification alleviates the problems of adverse selection that are present in noncertified sales. We find that nearly all of the difference in prices between certified and noncertified sales is attributable to the selection process.

The empirical strategy adopted allows us to use three approaches to test for adverse selection. The first approach examines the relationship between the residuals from participation and price equations by treating adverse selection as self-selection bias. In noncertified sales the data indicate horses that would receive unusually high prices in noncertified sales, given their observed attributes, are even more valuable in other options. In certified sales we find that self-selection does not affect the results.

We also find that seller characteristics play a role in noncertified sales but have no significant effect on either the probability that a horse is offered for sale or prices in certified sales. Using Lee's (1995) methodology to estimate opportunity costs, we find that over 87% of the difference in prices between certified and noncertified sales can be attributed to the selection process. Finally, we find that noncertified horses earn less per race than horses retained by their breeders.

Although our results provide compelling evidence that certification alleviates problems of adverse selection, we do not isolate the exact reasons for its success. Information on potential buyers' inspection expenditures in certified and noncertified sales and the effect these potential differences have on seller incentives to sell or retain horses might provide insight into the reasons for certification's effectiveness.

The techniques used here may be applied to other industries. For example, certification appears to play a major role in online markets, where it is generally impossible for potential buyers to inspect the quality of goods sold. Our results suggest that certification overcomes problems of adverse selection in a market where inspection is possible. It is reasonable to expect certification to be even more valuable in cases where buyers are unable to inspect the quality of goods offered for sale.

TABLE 1

Summary Statistics

 Full Sample Retained Noncertified

Certified 0.05 -- --
 (0.21) -- --
Noncertified 0.23 -- 1.00
 (0.42) -- --
Price -- -- 13,268
 -- -- (17,697)
Starter 0.69 0.65 0.82
 (0.46) (0.48) (0.38)
U.S. Earnings * 21,028 17,716 27,188
 (47,622) (45,094) (49,580)
U.S. Earnings per Race ** 966 798 1,388
 (2,158) (1,900) (2,567)
Racing Intensity 3.39 3.77 2.27
 (22.89) (25.32) (14.94)
Age in months 8.45 8.39 8.59
 (1.23) (1.23) (1.22)
Colt 0.50 0.50 0.52
 (0.50) (0.50) (0.50)
Sire Grade 1 0.24 0.17 0.38
 (0.42) (0.37) (0.49)
Dam Stakes Winner 0.10 0.07 0.15
 (0.30) (0.25) (0.36)
In (Earnings per Sire Offspring) 5.11 4.48 6.56
 (4.67) (4.63) (4.35)
Stakes-Winning Siblings 0.17 0.12 0.22
 (0.48) (0.43) (0.51)
Siblings 4.34 4.07 4.96
 (3.01) (2.86) (3.22)
Sire's Crop 23.29 18.34 33.74
 (20.22) (17.42) (20.73)
Certified State 0.47 0.39 0.62
 (0.50) (0.49) (0.49)
Sale State 0.82 0.79 0.91
 (0.38) (0.41) (0.28)
Unlisted Breeder 0.65 0.73 0.50
 (0.48) (0.45) (0.50)
Unlisted Sire 0.45 0.51 0.30
 (0.50) (0.50) (0.46)
Number observations 3,376 2,451 766

 Certified

Certified 1.00
 --
Noncertified --
 --
Price 93,437
 (105,155)
Starter 0.85
 (0.36)
U.S. Earnings * 48,475
 (66,417)
U.S. Earnings per Race ** 2,506
 (3,117)
Racing Intensity 2.89
 (12.73)
Age in months 8.69
 (1.16)
Colt 0.55
 (0.50)
Sire Grade 1 0.63
 (0.48)
Dam Stakes Winner 0.31
 (0.47)
In (Earnings per Sire Offspring) 7.95
 (4.11)
Stakes-Winning Siblings 0.56
 (0.85)
Siblings 5.35
 (3.52)
Sire's Crop 49.25
 (18.86)
Certified State 0.89
 (0.31)
Sale State 0.92
 (0.27)
Unlisted Breeder 0.25
 (0.43)
Unlisted Sire 0.21
 (0.41)
Number observations 159

Standard deviation in parentheses.

* Foreign earners excluded.

** Horses with zero starts receive value of zero in calculation.

TABLE 2

Price, Participation, and Earnings Regressions

 ln(Price) Regressions
Variable Noncertified Certified

Racing Intensity 0.002 -0.0005
 (0.76) (0.10)
Age in Months (a) 0.011 -0.028
 (1.43) (0.63)
Colt 0.033 0.418 ***
 (0.47) (3.46)
Sire Grade 1 0.131 0.168
 (1.42) (1.16)
Dam Stakes Winner 0.385 *** 0.486 ***
 (3.72) (3.51)
ln(Earnings per Sire Offspring) 0.277 *** 0.500 **
 (5.00) (2.01)
Stakes-Winning Siblings 0.632 *** 0.392 ***
 (7.02) (3.98)
Siblings -0.045 *** -0.035 *
 (2.84) (1.65)
Sire's Crop 0.010 *** 0.002
 (2.91) (0.40)
Inverse Mills ratio -1.312 *** -0.260
 (4.24) (0.65)
[[beta].sub.[lambda]s] -0.004 -0.0005
 ([[partial][lambda].sub.s]/
 [partial]RI)
 (1.62) (0.04)
Certified State -- --

Sale State -- --

Certified -- --

Noncertified -- --

Unlisted Breeder -0.030 0.093
 (0.33) (0.48)
Unlisted Sire 2.624 *** 4.80 **
 (4.96) (1.98)
Constant 7.234 *** 6.194 *
 (11.27) (1.77)
[R.sup.2] (log likelihood) 0.42 0.46
Number observations 766 159

 Multinomial Logit
Variable Noncertified Certified ln(U.S.) (b)

Racing Intensity -0.001 ** -0.0001 -0.002
 (2.23) (1.03) (1.00)
Age in Months (a) 0.01 ** 0.001 -0.045
 (2.35) (1.53) (1.05)
Colt 0.015 0.002 -0.093
 (0.99) (1.23) (0.88)
Sire Grade 1 0.067 *** 0.007 ** 0.110
 (3.44) (2.44) (0.78)
Dam Stakes Winner 0.03 0.005 0.316 *
 (1.32) (1.61) (1.67)
ln(Earnings per Sire Offspring) -0.009 0.012 *** 0.137 *
 (0.87) (3.84) (1.71)
Stakes-Winning Siblings -0.042 ** 0.002 0.133
 (2.53) (1.57) (1.09)
Siblings 0.013 *** 0.0004 -0.019
 (4.97) (1.14) (0.99)
Sire's Crop 0.005 *** 0.0004 *** 0.006 *
 (10.37) (4.41) (1.69)
Inverse Mills ratio -- -- --

[[beta].sub.[lambda]s] -- -- --
 ([[partial][lambda].sub.s]/
 [partial]RI)

Certified State 0.018 0.028 *** --
 (1.03) (3.93)
Sale State 0.089 *** -0.025 ** --
 (4.42) (1.75)
Certified -- -- -0.113
 (0.41)
Noncertified -- -- 0.290 **
 (2.15)
Unlisted Breeder -0.073 *** -0.01 *** -0.339 ***
 (4.23) (3.08) (2.73)
Unlisted Sire -0.239 *** 0.851 *** 1.413 **
 (4.02) (5.85) (1.91)
Constant -- - 6.462 ***
 (7.77)
[R.sup.2] (log likelihood) -1,968 -1,968 -6,393
Number observations 3,376 3,376 3,183

Absolute value of t or z statistics in parentheses.

(a)Age in months for logit and earnings regressions; age at sale in
price regressions.

(b)Foreign runners excluded. Corrected for self-selection of whether
horse started a race.

* Significant at 0.10 level.

** Significant at 0.05 level.

*** Significant at 0.01 level.

(1.) The literature has examined the role of information asymmetries in a variety of markets. Examples include credit markets as in Ausubel (2000), insurance markets examined by Puelz and Snow (1994) or Chiappori and Salanie (2000), automobile markets studied by Bond (1982) or Genesove (1993), and the market for initial public offerings as in Booth and Smith (1986) and Gompers and Lerner (1999).

(2.) See the discussions in Heckman (1979) and Lee (1983). Because breeders decide between retaining a horse, selling it in a noncertified sale, and nominating it to a certified sale, we use Lee's (1983) correction for sample-selection bias in polychotomous-choice models.

(3.) Matthews and Postlewaite (1985) examine testing and disclosure requirements. Heinkel (1981) and Mason and Strebenz (1994) examine imperfect tests of product quality. Albano and Lizzeri (2001) treat the amount of information an agent discloses as a strategic variable. Biglaiser and Friedman (1999) show that when the first best is not attainable, a second best characterized by intervals of high and low-quality goods being sold exists. Spulber (1999) provides an excellent overview of the literature regarding intermediaries.

(4.) The Thoroughbred Owners and Breeders Association "grade" high-quality races. In 2000 only 96 of the thousands of races run received the highest grade (grade I). In our data only 4 of the 3,183 horses that ran in the United States won a grade I race, and 1,161 did not earn any money at the track.

(5.) In 1994, Keeneland Sales Company had approximately 3,500 yearlings pass through its sales ring in its September sale, and no buyer purchased more than 30 horses at this sale. This sale accounted for approximately one-third of young thoroughbreds sold and 10% of all horses born in the United States in 1993.

(6.) Though the particulars of each auction house's certification process may differ slightly all follow the same basic procedure.

(7.) Keeneland reports that approximately 50% of the horses nominated to its certified sales are eliminated based on the initial ranking. This information was obtained through interviews with Rogers Beasley, Keeneland's sales director. We are unable to report specifics for other auctions in our sample.

(8.) Panels consist of experienced market experts, which may include a licensed veterinarian. Each panel may evaluate hundreds of young thoroughbreds each year.

(9.) When selling goods from single producers, intermediaries may have the incentive to collude with producers that misrepresent quality.

(10.) Viscusi (1978), Metzger (1983), Mason (1986) and Shapiro (1986) also examine minimum quality standards.

(11.) Such tests include x-ray and sonograms. Many buyers hire agents that inspect the quality of horses and offer recommendations about a horse's value.

(12.) This approach is consistent with Chiappori and salanie (2000), who examine the relationship between errors from selection and performance equations.

(13.) In our data there are a total of 53 sales, 6 of which are certified sales. There are fewer than 10 observations for many of these sales.

(14.) Gamrat and Sauer (2000) show that the average thoroughbred owner receives a negative return from racing horses, concluding that racehorse owners receive nonpecuniary benefits from racing. We assume that breeders maximize a well-behaved utility function that is a function of a horse's quality, a Hicksian composite commodity, breeder characteristics, and other exogenous factors, subject to a budget constraint. This approach does not preclude the possibility that breeders sell horses because they are capacity constrained as assumed by Genesove (1993).

(15.) This includes both racetrack earnings and a horse's residual value in breeding operations.

(16.) Because prices are determined at auction, in our data the choice of selling is based on expected prices. For the choice of nominating a horse to a certified sale breeders must account for the probability a horse is certified.

(17.) We are unable to identify horse that are nominated to a certified sale but rejected. These horse are either retained or sold in noncertified sales and are coded accordingly.

(18.) See Lee (1995) for details demonstrating that "normalization" of the probabilities obtained from the multinomial logit allow implementation of a correction for self-selection bias similar to Heckman's (1979).

(19.) Following Lee (1983) we correct the variance-covariance matrix for heteroscedasticity using Heckman's two-step procedure as shown by Greene (1995).

(20.) As shown by Lee (1983) the coefficient on the Mills ratio is given by the covariance between the transformed errors of the participation equation and the errors from the price equation.

(21.) Reimers (1985) discusses interpretation of the coefficient on the inverse Mills ratio in a labor-market context.

(22.) Chezum and Wimmer (1997) argue that in addition to physical constraints, as the size of a breeder's racing operating grows it becomes increasingly difficult to monitor trainers and the performance of racehorses. In a utility-maximizing framework, it is optimal for breeders to sell their lowest-quality horses first when faced with a capacity constraint and asymmetric information.

(23.) Because this term is nonlinear we use a standard bootstrap program to calculate its standard error, evaluated at the sample mean. We also include RI in the price equation directly. The coefficient on RI in the price equation captures any direct correlation between the expected quality of horses, included retained horses, and RI.

(24.) It should be noted that our findings might be driven in part by information that is observed by market participants but not measured in our data. This is particularly true for the results related to the inverse Mills ratio. Consider the case in which information is perfect and symmetrical. In this setting, capacity-constrained sellers may sell low-quality horses because they place a higher value on quality than buyers and use sales to cull their bad horses. If certification simply sorts horses based on perfect information and place high-quality horses in certified sales, sellers culling horses would not participate in certified sales. If this is the case, we expect to find that an increase in RI reduces the probability of inclusion in certified sales by more than it does in noncertified sales. Our second test, therefore, provides evidence that is less sensitive to missing-variable problems.

(25.) In contrast to Lee (1995), who enters a negative inverse Mills ratio, we enter the inverse Mill ratio directly. Differences in interpretation follow.

(26.) As discussed by Gamrat and Sauer (2000), racetrack earnings are an imperfect measure of a horse's quality because horses also have value in breeding operations. High-quality horses are likely to retired sooner to begin breeding and earnings as a measure of quality are biased downward. We use earning per race to overcome this problem.

(27.) The Jockey Club reports that 36,455 horses born in 1993 were registered, of which 33,174 were born in the United States.

(28.) Horses are listed in alphabetic order by the name of their dam (mother), which is independent of quality.

(29.) In the thoroughbred industry, thoroughbreds are classified as yearlings on January 1 of the year following their birth. The analysis includes a small number of horses that were sold in November of their first year.

(30.) For horses bred, the breeder may or may not be the owner of the horse at the time the horse races.

(31.) We are unable to gather information on stud fee, which is a market measure of a sire's quality, for a large number of sires in our sample.

(32.) We exclude horses that ran in foreign countries because of fundamental differences in the amount horses are paid for winning races across countries. Approximately 6% of the horses in our sample ran in foreign countries.

(33.) Any bias resulting from our inability to identify horses from these two groups works against a finding of adverse selection in noncertified sales. In the earningsper-race equations we employ Heckman's correction for self-selection to control for this, using a qualitative variable that indicates whether a horse's sire started a race to identify the equation. Because all sires in certified sales started a race, we are unable to use this variable in multinomial-logit model.

(34.) In additional regressions, not reported, we treat the choices of selling in noncertified and certified sales as independent and use the classic Heckman technique. These results are consistent with results reported here.

(35.) Because of nonlinearities, standard errors are obtained by running a standard bootstrapping program with 500 iterations.

(36.) We perform a Chow test on the data and reject the null that the coefficients on all independent variables in the certified and noncertified sales are jointly equal.

(37.) Though the marginal effect and its underlying coefficient of Racing Intensity is not statistically different from zero, we cannot reject the null that the underlying coefficients on Racing Intensity in the certified and noncertified equations are equal. As discussed in note 24, a test for whether or not the results are driven by information available to the market but not the econometrician is a finding that the Racing Intensity has a larger impact (negative) in certified sales than in noncertified sales. The data reject this hypothesis.

(38.) For each of the independent variables in the multinomial logits we tested the null hypothesis that the underlying coefficients in the certified and noncertified equations are equal. We reject the null hypothesis for Log Sire Earnings Per Offspring, Stakes-Winning Silblings, Sire's Crop, Certified State, Sale State, Unlisted Breeder, and Unlisted Sire. We also tested the null hypothesis that the coefficients are jointly equal and reject the null at 0.001 level of significance.

(39.) To calculate these numbers we estimated the expected prices for every horse in the sample and took a simple average. The magnitude of the effect is similar for different subsets of the data.

(40.) This result is altered slightly by looking at different subsets of the data. Using only the average expected prices from noncertified sales we find that E[[P.sub.C]\X, Z] - E[[P.sub.NC]\X, Z, NC = 1] = $81,932-$13,984 = $67,948, but the expected price for a random horse in noncertified sales equals $108,049. Obviously, differences in the value buyers place on observed attributes differs across the sales and result in a reversal in order of the expected price in certified and noncertified sales.

(41.) Because approximately 31% of the horses in our sample never raced, we correct the regression for self-selection bias using whether or not the sire ever raced to identify the selection equation. In the participation equation, which is not reported here, we find that horses sold in either certified or noncertified sales are more likely to race than horses that were not sold. This result suggests that a portion of the horses in our sample were bred for reasons other than racing.

REFERENCES

Akerlof, G. A. "The Market for "Lemons": Quality Uncertainty and the Market Mechanism." Quarterly Journal of Economics, 84(3), 1970, 488-500.

Albano, G. L., and A. Lizzeri. "Strategic Certification and Provision of Quality." International Economic Review, 42(1), 2001, 267-83.

Ausubel, L. M. "Adverse Selection in the Credit Card Market." Mimeo, University of Maryland, 2000.

Barzel, Y. "Measurement Cost and the Organization of Markets." Journal of Law and Economics, 25(1), 1982, 27-48.

Biglaiser, G. "Middlemen as Experts." Rand Journal of Economics, 24(2), 1993, 212-23.

Biglaiser, G., and J. W. Friedman. "Middlemen as Guarantors of Quality." International Journal of Industrial Organization, 12(4), 1994, 509-31.

-----. "Adverse Selection with Competitive Inspection." Journal of Economics and Management Strategy, 8(1), 1999, 1-32.

Bloodstock Research Information Systems, Inc. American Produce Records. Lexington, 2000.

Bond, E. W. "A Direct Test of the "Lemons" Model: The Market for Used Pickup Trucks." American Economic Review, 72(4), 1982, 836-40.

Booth, J., and R. Smith. "Capital Raising, Underwriting and the Certification Hypothesis." Journal of Financial Economics, 15(1-2), 1986, 261-81.

Chezum, B., and B. S. Wimmer. "Roses or Lemons: Adverse Selection in the Market for Thoroughbred Yearlings." Review of Economics and Statistics, 79(3), 1997, 521-26.

Chiappori, P., and B. Salanie. 2000. "Testing for Asymmetric Information in Insurance Markets." Journal of Political Economy, 108(1), 2000, 56-78.

Daily Racing Form. The American Racing Manual: 1993. Lexington, 1993.

-----. The American Racing Manual: 1994. Lexington, 1994.

Finnochio, E. J. "The Significance of Birthrank, Maternal Age and Parity on Race Performance in Thoroughbreds." Unpublished manuscript, 1995.

Gamrat, F. A., and R. D. Sauer. "The Utility of Sport and Returns to Ownership: Evidence from the Thoroughbred Market." Journal of Sports Economics, 1(3), 2000, 219-35.

Genesove, D. "Adverse Selection in the Wholesale Used Car Market." Journal of Political Economy, 101(4), 1993, 644-65.

Gompers, P., and J. Lerner. "Conflict of Interest in the Issuance of Public Securities: Evidence from Venture Capital." Journal of Law and Economics, 42(1), 1999, 1-28.

Greene, W. H. Econometric Analysis. Upper Saddle River, NJ: Prentice Hall, 1997.

-----. LIMDEP, Version 7.0 User's Manual, Econometric Software, Inc. Bellport, NY: Econometric Software, 1995.

Heckman, J. J. "Sample Selection as a Specification Error." Econometrica, 47(1), 1979, 153-62.

Heinkel, R. "Uncertain Product Quality: The Market for Lemons with an Imperfect Testing Technology." Bell Journal of Economics, 12(2), 1981, 625-36.

Lee, L. "Generalized Econometric Models with Selectivity." Econometrica, 51(2), 1983, 507-12.

-----. "The Computation of Opportunity Costs in Polychotomous Choice Models with Selectivity." Review of Economics and Statistics, 77(3), 1995, 423-35.

Leland, H. E. "Quacks, Lemons, and Licensing: The Theory of Minimum Quality Standards." Journal of Political Economy, 87(6), 1979, 1328-46.

Mason, C. E. "Cherries, Lemons, and the FTC, Revisited [The Market for 'Lemons'][Cherries, Lemons, and the FTC: Minimum Quality Standards in the Retail Used Automobile Industry]." Economic Inquiry, 24(2), 1986, 363-65.

Mason, C. F., and F. P Sterbenz. "Imperfect Product Testing and Market Size." International Economic Review, 35(1), 1994, 61-86.

Matthews, S., and A. Postlewaite. "Quality Testing and Disclosure." Rand Journal of Economics, 16(3), 1985, 328-40.

Metzger, M. R. "Cherries, Lemons, and the FTC: Minimum Quality Standards in the Retail Used Automobile Industry." Economic Inquiry, 21(1), 1983, 129-39.

Puelz, R., and A. Snow. "Evidence on Adverse Selection: Equilibrium Signaling and Cross-Subsidization in the Insurance Market." Journal of Political Economy, 102(2), 1994, 236-57.

Reimers, C. W. "A Comparative Analysis of the Wages of Hispanics, Blacks, and Non-Hispanc Whites," in Hispanics in the U.S. Economy, edited by G. J. Borjas and M. Tienda. San Diego, CA: Academic Press, 1985, 27-56.

Shapiro, C. "Investment, Moral Hazard, and Occupational Licensing." Review of Economic Studies 53(5), 1986, 843-62.

Spulber, D. F. Market Microstructure: Intermediaries and the Theory of the Firm. Cambridge: Cambridge University Press, 1999.

Stigler, G. J. "The Economics of Information." Journal of Political Economy, 69(3), 1961, 213-25.

The Jockey Club. American Stud Book, Foals of 1993. New York, 1995.

Trost, R. P., and L. Lee. "Technical Training and Earnings: A Polychotomous Choice Model with Selectivity." Review of Economics and Statistics, 66(1), 1984, 151-56.

Viscusi, W. K. "A Note on 'Lemons' Markets with Quality Certification." Bell Journal of Economics, 9(1), 1978, 277-79.

BRADLEY S. WIMMER and BRIAN CHEZUM *

* We thank Alan Schlottmann and two anonymous referees for helpful comments. We also thank Dotti Britt for competent research assistance. All mistakes are ours alone.

Wimmer: Assistant Professor, University of Nevada Las Vegas, Las Vegas, NV 89154-6005. Phone 1-702-895-3018, Fax 1-702-895-1354, E-mail wimmer@ccmail. nevada.edu

Chezum: Florence Irene Eggleston Associate Professor, St. Lawrence University, Canton, NY 13617. Phone 1-315-229-5426, Fax 1-315-229-5819, E-mail chezum@stlawu.edu