Breeding racehorses: what price good genes?

Abstract

Horse racing is a multi-million pound industry, in which genetic information is increasingly used to optimize breeding programmes. To maximize the probability of producing a successful offspring, the owner of a mare should mate her with a high-quality stallion. However, stallions with big reputations command higher stud fees and paying these is only a sensible strategy if, (i) there is a genetic variation for success on the racecourse and (ii) stud fees are an honest signal of a stallion's genetic quality. Using data on thoroughbred racehorses, and lifetime earnings from prize money (LE) as a measure of success, we performed quantitative genetic analyses within an animal model framework to test these two conditions. Although LE is heritable (V_A=0.299±0.108, Pr=0.002), there is no genetic variance for stud fee and the genetic correlation between traits is therefore zero. This result is supported by an absence of any relationship between stud fees for currently active stallions and the predicted LE for their (hypothetical) offspring. Thus, while there are good genes to be bought, a stallion's fees are not an honest signal of his genetic quality and are a poor predictor of a foal's prize winning potential.

1. Introduction

Thoroughbred horse racing is a multi-million pound international industry, in which there is an increasing emphasis on using genetic tools to optimize breeding programmes. Both quantitative and molecular approaches have revealed genetic variation for key racing performance traits (Gaffney & Cunningham 1988; Hill 1988; Harrison & Turrion-Gomez 2006), and accurately assessing the genetic potential of candidate parents may therefore be critical to breeding successful racehorses (Bailey 1998). In particular, the owner of a thoroughbred mare should mate her with a high-quality stallion in order to maximize the probability of producing successful progeny.

However, a breeder must also pay stud fees which increase with a stallion's reputation, a reputation that might be based on his own racing career, or on the performances of his progeny to date (or both). Thus, by paying higher stud fees, breeders may assume that they are effectively buying better genes and thereby increasing the prize-winning potential of the resultant foal. However, this is only true if the stallion's reputation, and hence nomination fee, is an honest signal of his genetic quality. This system has strong parallels in evolutionary biology, where theory predicts that honest signals of male genetic quality will be used by females to determine mate choice and influence levels of reproductive investment (e.g. Fisher 1958; Kokko et al. 2002; Maynard Smith & Harper 2003). Here, however, investment and mate choice are performed by proxy (i.e. by the breeder rather than the mare), with a view to maximizing financial returns rather than evolutionary fitness. In this context, paying higher fees for a stallion with a bigger reputation may not make financial sense if fees are not an honest signal of genetic quality.

Here we perform quantitative genetic analyses on a pedigreed dataset of thoroughbred racehorses to address two questions. Firstly, we test whether there is genetic variation for success on the racecourse by analysing data on lifetime prize money earnings. Secondly, we ask whether stud fees are a useful indicator of a stallion's genetic quality and hence its offspring's prize-winning potential.

2. Material and methods

(a) Data sources

Data relating to 554 currently (or recently) active stallions standing at stud in the UK and USA were collated from freely available sources at www.the-racehorse.com, http://www.britishhorseracing.com/ and http://breeding.bloodhorse.com. For each stallion we recorded stud farm, state (or country) and stallion nomination fee (SNF). For some stallions, fee records were available from multiple seasons (between 2001 and 2007). In such cases, we averaged over records for simplicity. Analyses using multiple records per individual were also performed, but they yielded quantitatively similar results and are not shown.

Lifetime prize money earned (LE) was recorded for each stallion, and also for each of his ancestors (both male and female) to a depth of four generations, from http://www.pedigreequery.com/. For each horse, the number of races run, and year and country of foaling were obtained from the same source. In total, we thus compiled a pedigreed database of 4476 horses foaled between 1922 and 2003, with stud fees for 554 active stallions and lifetime performance statistics for a much larger set of 2500 horses. All lifetime earnings and stud fees were converted to US dollars where reported in other currencies (based on exchange rates in the year of birth) and were log_e transformed prior to analysis. Data sources provide no guarantee as to the accuracy of the data, and while errors in either phenotypic values or pedigree links will cause downward bias in estimated genetic variance, this effect is expected to be minimal.

(b) Quantitative genetic analyses

Phenotypic variance for the traits of LE and SNF was partitioned into genetic and environmental variance components using univariate animal models (Kruuk 2004) solved by restricted maximum likelihood using the program ASReml (VSN International). For each trait y, the phenotype of individual i (y_i) was modelled as

$$y_i=\mu +\text{fixed}\text{effects}+a_i+e_i,$$ (2.1)

where μ is the overall population mean, a_i is the additive genetic effect on individual i assumed to be normally distributed with mean of zero and variance of V_A (the additive genetic variance) and e_i is a residual resulting from environmental effects with zero mean and variance V_R. The genetic variance V_A is estimated from the variance–covariance matrix of additive genetic effects, which is equal to $\mathit{A}{\sigma }_{\text{A}}^2$, where A is the additive numerator relationship matrix containing the individual elements A_ij=2Θ_ij, and Θ_ij is the coefficient of coancestry between individuals i and j (obtained from the pedigree structure).

In addition to the population mean μ, further fixed effects were included to account for known influences on phenotype. Statistical significance of fixed effects was assessed using conditional Wald F statistics. For LE, fixed effects included were the number of races started, sex, year of birth (modelled as a quadratic) and country of birth (as a factor). For SNF, state (or country if outside the USA) was included as a fixed effect while the stud farm was included as an additional random effect (with a corresponding variance component, V_STUD consequently partitioned). For both traits additional random effects of sire, dam and grandsire (both maternal and paternal) were also fitted and their significance tested using likelihood ratio tests. These effects were tested both to protect against upward bias of V_A estimates (Wilson et al. 2005) and to examine the possibility of traits showing particular maternal inheritance (Harrison & Turrion-Gomez 2006). Non-significant parental and grandparental effects were dropped before narrow sense heritability was estimated as the ratio of V_A to total phenotypic variance V_P (calculated as the sum of variance components). A bivariate model was then used to estimate the genetic covariance and correlation (r_G) between traits.

Finally, to further scrutinize the potential association between genetic merit for LE and SNF, we generated predicted lifetime earnings, PLE (on the natural log scale), for a hypothetical progeny (p) of each currently active stallion (i) as

$${\text{PLE}}_p={\hat{\mu }}_{\text{F}}\hspace{0.17em}+\hspace{0.17em}0.5{\hat{a}}_i,$$ (2.2)

where ${\hat{\mu }}_{\text{F}}$ is the predicted fixed effect mean, which (for arbitrary illustrative purposes) was evaluated for female offspring foaled in the USA in 2002 which were all assumed to start 13 races; and ${\hat{a}}_i$ is the predicted genetic merit of the sire (i.e. the best linear unbiased predictor of a_i) which is multiplied by 0.5 to account for Mendelian transmission for sire to daughter. Values of PLE_p were then exponentiated to yield predictions on the absolute (dollar) scale and regressed on stud nomination fees (dollar scale) of active sires.

3. Results

Animal model analysis indicated a heritable component of variation for lifetime prize earnings (LE; table 1). Additive genetic variance was found to be statistically significant based on comparison to a reduced model (V_A=0.299±0.108; Χ₁²=10.02, Pr=0.002), with a corresponding heritability, h² of 0.095±0.034. Inclusion of parental and grandparental effects did not result in significantly improved likelihoods and we therefore do not present results from these more complex models.

Table 1.

Univariate animal model analysis of log_e transformed lifetime prize earnings (LE) in thoroughbred racehorses. (Phenotypic variance (V_P) was partitioned into additive genetic (V_A) and residual environmental (V_R) components. The corresponding ratios of these components to phenotypic variance are indicated. Fixed effects included in the model are also shown with statistical significance assessed using conditional Wald F statistics. Note that the coefficient for sex is the expected effect in males relative to females.)

component	variance (s.e.m.)			ratio to VP (s.e.m.)
total phenotypic (VP)	3.130 (0.093)			1 (—)
additive genetic (VA)	0.299 (0.108)			0.095 (0.034)
residual (VR)	2.832 (0.125)			0.905 (0.034)

	coefficient (s.e.m.)	numerator d.f.	denominator d.f.	F	Pr
fixed effects
μ	1	1	109.6	33 440.6	<0.001
no. races started	0.050 (0.003)	1	2144.4	482.53	<0.001
sex	2.141 (0.080)	1	2286.7	884.19	<0.001
year of birth	0.052 (0.003)	1	611.4	352.55	<0.001
year of birth2	0.004 (0.002)	1	2091.2	96.33	<0.001
country of birth	—	17	1967.2	2.74	<0.001

In contrast to LE, there was no evidence of additive genetic variance for SNF. V_A was estimated as −0.015±0.052, and fixed at zero when the model was constrained to biologically meaningful parameter space (i.e. variance components greater than or equal to zero). Substantial differences were found in stallion nomination fees among farms (univariate analysis of SNF; V_STUD=0.4176±0.077). This effect accounted for 44% of the phenotypic variance in SNF and was statistically significant based on a likelihood ratio test comparison with a reduced model (Χ₁²=130, Pr<0.001).

The phenotypic correlation between the log_e-transformed traits was positive but marginally non-significant (r_LE.SNF=0.074, Pr=0.074), while in the absence of additive genetic variance for SNF (a conclusion also supported by the bivariate animal model; results not shown) the genetic correlation between LE and SNF is defined as zero. Further support for the absence of a genetic relationship between LE and SNF is provided by model predictions of progeny lifetime earnings for currently active stallions (figure 1). Among stallions of varying genetic merit for LE, predicted lifetime progeny earnings for hypothetical female offspring varied from approximately $57 500 to $140 000. However, the slope of a linear regression of predicted lifetime earnings on SNF (performed on the dollar scale), though positive, was small (0.020±0.011) and insignificant (F_1,438=2.98, Pr=0.085). It should also be noted that since predicted lifetime progeny earnings incorporate estimates of genetic merit they are not independent among relatives, and statistical testing in this way is also inherently anti-conservative (Hadfield in press).

Figure 1.

Predicted progeny lifetime prize earnings plotted against stud nomination fees. Each data point represents the predicted return from the progeny of a recently or currently active stallion at stud in USA or UK. The linear regression line has a slope of 0.02 indicating an expected return of two cents in prize money for each additional dollar spent on fees. (Note that maximum SNF_max was actually 500 000, but we have truncated the axis for clarity).

4. Discussion

Our analysis shows that there is heritable variation for lifetime prize money earned. While environmental effects (including, for example, training, choice of races entered, jockeys employed and stochastic injury events) clearly have a much greater influence on lifetime earnings (accounting for 91.5% of the variance), LE is a heritable trait that could respond to appropriate selection. Thus, there is variation in genetic quality, and it should be possible for a breeder to pick a stallion for his good genes.

While there are good genes to be bought, it does not appear to be the case that you get what you pay for. Rather than having any underlying genetic basis, our analyses show that the phenotypic association between fees and lifetime earnings arises from environmental, not genetic, effects. Thus, it seems probable that breeders who can provide the best environment and expertise for training race horses are also likely to be those best able to afford the highest stud fees. However, there is certainly no support for the hypothesized positive genetic correlation between stallion nomination fees and lifetime prize earnings. Although power to directly estimate genetic parameters for SNF is limited (discussed further below), our conclusion is also supported by the regression of predicted progeny earnings on sire stud fee. While a regression line with slope of 1 would mean an expected financial return that was equal to the investment, here the slope is close to zero (with an expected return of $0.02 for each additional dollar spent on fees) and not significant.

It is important to note that prize money is by no means the only route to financial return in the racing industry (e.g. stallions may earn considerably more at stud than on the track) and it would also be wrong to suggest that making money is the sole reason for owning or breeding racehorses. Consequently, it could be readily argued that maximizing expected lifetime prize earnings may not be the most important objective for a racehorse breeder. Nevertheless, in the absence of genetic variance for stallion nomination fees, it is not possible for SNF to show genetic covariance with other performance traits that might be more direct targets of selection. That is to say, the fees are not expected to correlate well with genetic merit for any other trait.

As a partial caveat to this conclusion, we acknowledge that sample size is relatively small for SNF and that this may limit power to detect low, but non-zero, levels of V_A (at least by comparison with LE). Nevertheless, based on the unconstrained animal model we determined an upper (95%) confidence limit of just 0.09 for V_A (approx. 8% of the phenotypic variance as estimated using a model containing fixed effects only). Furthermore, our conclusion is based on available stud fee data from 2001 to 2007 and it is possible that V_A for SNF varies over time. For example, genotype-by-environment interactions are ubiquitous in nature and mean that genetic (co)variances change over time in temporally heterogeneous environments (Wilson et al. 2006). To the extent that the ‘economic environment’ in which stud fees are determined may also be variable, there is no reason to preclude similar effects here.

In summary, our results show that genetic variance exists for lifetime prize earnings as well as more specific performance traits measured on the track (e.g. Gaffney & Cunningham 1988). However, if the goal is to maximize lifetime prize winnings, then it seems clear that stud fees are not an honest signal of a stallion's genetic quality.