Inequality of maximum a posteriori estimators with equivalent sire and animal models for threshold traits

Inegalite entre des estimateurs du maximum a posteriori avec des modeles equivalents pere et animal pour des caracteres a seuil. Il est demontre que l'utilisation des modeles equivalents pere et animal, pour decrire la variable sous-jacente d'un caractere a seuil, conduit a differents estimateurs du maximum a posteriori. Dans les exemples simples analyses dans cette etude, en comparant l'utilisation de ces 2 modeles on constate que les estimateurs du maximum a posteriori tendent a se rapprocher de 0 lorsqu'on utilise le modele animal. Les differences entre les 2 estimateurs sont particulierement notables quand l'heritabilite sur l'echelle sous-jacente est elevee, et elles augmentent avec l'heritabilite. De plus, il est aussi montre que meme le classement des reproducteurs peut etre different selon l'estimateur utilise comme critere de classement.


INTRODUCTION
New procedures based on the threshold concept have recently been developed for the genetic analysis of ordered categorical data in animal breeding. The methods introduced by Gianola and Foulley (1983), Stiratelli et al (1984) and Harville and Mee (1984) are essentially the same and were derived mainly in a Bayesian framework. With these methods, location parameters in the underlying scale are estimated by maximizing a joint posterior density. The estimators are therefore designated as maximum a posteriori estimators. Studies on the properties of maximum a posteriori estimators and comparisons with the estimators and predictors of linear model methods have previously been based on sire models to describe the variable on the underlying scale (eg, Meijering and Gianola, 1985;H6schele, 1988;Weller et al, 1988;Renand et al, 1990;Mayer, 1991). For continuously distributed traits the use of an animal model is the state of the art for the genetic evaluation of breeding animals. A logical step would be to use an animal model as well to describe the underlying variable in genetic evaluation with threshold traits. In the present paper some properties of maximum a posteriori estimators are studied in the context of applying an animal model to describe the underlying variable of threshold traits.

Example data set
Consider the following simple data structure (2 sires, each with 1 progeny) for a dichotomous character with the 2 realizations M o (0) and M, (1): Progeny 1 exhibits the realization M o and progeny 2 the realization M l . Assume that we want to estimate the breeding values of the sires for the trait M.

Maximum a posteriori estimation
To estimate the breeding values of the sires, the threshold concept is used and we follow the Bayesian approach along the lines introduced by Gianola and Foulley (1983), Stiratelli et al (1984), and Harville and Mee (1984). Let the data be arranged as a 2 x 2 contingency table, where the 2 rows represent, as usual, the sire effects, and the 2 columns represent the realizations M o and M i . The elements of the contingency table in the ith row and jth column are designated as n ij . The columns indicate mutually exclusive and exhaustive ordered categories of response.
The rth experimental unit in row i is characterized by an underlying continuously distributed variable, which is rendered discrete through a fixed threshold.
Let us assume that the variance components in the description of the underlying variable are known, at least to proportionality. Using a 'flat' prior for the threshold (t), so as to mimic the traditional mixed model analysis, and assuming that the sire effects on the underlying scale (u l , u 2 ) a priori are normally, identically and independently distributed with mean 0 and variance Q u, the joint posterior density has the form ( C1 = constant): Using a probit-transformation and a purely additive genetic model for the underlying variable (4l: standardized normal cumulative density function): The residual standard deviation ( Qe ) is taken as the unit of measurement on the underlying scale, so that Qe = 1. The maximum a posteriori estimators are the values which maximize the joint posterior density g(/3). This is equivalent to the maximation of the log-posterior density, which can be done more easily. The first derivatives of the log-posterior density with respect to t and u i are not linear functions of these parameters and must be solved iteratively using numerical techniques such as Newton-Raphson or Fisher's scoring procedure. The Newton-Raphson algorithm leads to the following iterative scheme: where (0: standardized normal probability density function): The joint posterior density g((3) is symmetric in the sense that g(t = z i , U1 = X2 , U2 = .r3!!,!,!) = 9 (t = -X1 , U1 = -X3,U2 = -!2!!,!,!). From this expression it follows directly that the posterior density has its maximum where t = 0 and U1 + U2 = 0. Making use of this result, the iterative equation [2] reduces to: When using the Fisher-scoring algorithm q i in the above iterative scheme, equation [3] has to be replaced by: The estimated breeding values of the sires equal 2u !k) at convergence. For compatibility reasons with the maximum a posteriori estimators in the following section, which are based on an animal model to describe the underlying variable and where the environmental standard deviation ( QE ) is taken as the unit of measurement, the estimator in [3] may also be expressed in units of QE by multiplying u!k) by [(1 -4 c )/(1 -C )]-1 / 2 or (1 -3<!)'!, respectively, where c is the intraclass correlation Q u/(1+Qu). Moreover, it is convenient to express the maximum a posteriori estimators in terms of the phenotypic standard deviation ap = ( U2 + !e)1!2 by multiplying the estimator from [3] by [1/(l -c )]-l / 2 or ( a2 + 1)-1!2, respectively.
Use of an equivalent animal model Up to now we have dealt with a sire model to describe the underlying variable for the threshold trait. Let the rows of the contingency table now represent the progenies in which the observations are made and we use an equivalent animal model to describe the basis variable and to estimate the breeding values of the sires. For that purpose let v' = [!1!2 !3 !4], where v l and v 2 represent the additive genetic values on the underlying scale of the progenies 1 and 2, and v 3 and v 4 represent the additive genetic effects of their sires 1 and 2.
Under these conditions v is a priori normally distributed with mean 0 and variance-covariance matrix Acr 2, where A is the numerator relationship matrix and a2 the additive genetic variance on the underlying scale. For the example data set the relationship matrix has the following form: The likelihood function invariably follows a product binomial distribution. So the joint posterior density is now: with and the environmental variance (!E) as the unit of measurement on the underlying scale.
For the joint posterior density [5] the Newton-Raphson algorithm leads after a little algebra to the following iterative equation system: where: It can be easily seen that for vi k ) and v2!! respectively, this equation system is formally very similar to the iterative equation (2!. Further, it shows the result that V i -Vi /2 and V4 2 /2, ie the estimated breeding values of the sires equal half of the estimated genetic effects of their progenies. To express the maximum a posteriori estimators in units of the phenotypic standard deviation they have to be multiplied, as in the case of the sire model by [1/(1 -c)!-1!2 or (ua + 1 )-1 / 2 , respectively.
More complex data structure A more complex data structure, viz the hypothetical data set already used by Gianola and Foulley (1983) to illustrate the maximum a posteriori estimation procedure, was employed to demonstrate the differences between the estimators based on either a sire or an animal model. The data consist of calving ease scores from 28 male and female calves born in 2 herd-years from heifers and cows mated to 4 sires. Calving ease was scored as a response in 1 of 3 ordered categories (normal birth, slight difficulty, extreme difficulty). The data are arranged into a contingency were postulated. For the maximum a posteriori estimation of the location parameters of the underlying variable, the computer program package TMMCAT (Mayer, 1991(Mayer, , 1994a) was used.
Different sire rankings Because of the inequality of estimated breeding values stemming from equivalent models to describe the underlying variable, there arose the question as to whether even the ranking of the sires may be different depending upon the model type chosen. To examine this question, data for 3 sires were systematically generated and analyzed. The number of observations n ij of sire i (i = 1,2,3) in response category j (j = 1, 2) was varied from 0 to 8. A simple random one-way model (sire model, animal model) for the underlying variable and a heritability value of 0.60 was postulated. Again for the analyses, the computer program TMMCAT was used.

RESULTS
Example data set Figure 1 shows the relationship between the estimated breeding value (EBV) of sire 1 and heritability for each of the models. The upper graph represents the estimates taking the environmental standard deviation as the unit of measurement, while in the lower graph the estimates are expressed in units of the phenotypic standard deviation.
It is obvious that the EBV is different whether a sire model or an equivalent animal model is used to describe the underlying variable. The difference between the EBVs is noticeable at a heritability of about 0.60 and increases with increasing heritability. With the phenotypic standard deviation as the unit of measurement, the EBV based on a sire model shows an almost linear relationship with the heritability even if with increasing heritability the slope slightly decreases. With a higher intraclass correlation on the underlying scale as with the animal model, this phenomenon can have a drastic impact. Thus, with the animal model, the EBV exhibits a maximum at a heritability value of about 0.81 and decreases rapidly thereafter.
Hypothetical data set by Gianola and Foulley  Moreover, table II contains the EBVs of the sires, which in the case of the sire model, for reasons of comparison, were calculated as twice the estimated sire effects.
For the heritability value of 0.20 the estimators from the sire model and the animal model are quite similar. The absolute estimates from the animal model are a little smaller, with the exception of the estimated distance between the thresholds. The greatest relative difference is shown by the contrast between cow effect and heifer effect. When a heritability of 0.60 is applied, the differences between the 2 model types are more substantial. Again, the absolute estimates from the animal model are smaller, only the estimated distance between the thresholds is greater and the greatest relative difference is found for the contrast between cow effect and heifer effect.
Different sire rankings Table III shows 3 data structures for which different sire rankings under the different model types (sire model and animal model) were identified.
The estimates based on the animal model are again smaller than the estimates from the sire model. The different rankings under the 2 model types are interesting. Under the sire model the EBVs of the sires 3 are clearly greater than the values of the sires 2, whereas under an animal model it is the other way around and the EBVs of the sires 3 are clearly smaller than the EBVs of the sires 2. Other data structures with different sire rankings were identified and table III shows only a few examples. It might be expected that with higher heritability values and more complex data designs this topsy-turvy situation would be even more marked.

DISCUSSION
Studies on the properties of maximum a posteriori estimators for threshold traits have been based on sire models to describe the underlying variable. Interest has mainly focused on comparisons of maximum a posteriori estimators with the estimators and predictors of linear model methods. As regards the EBVs, or more exactly sire evaluation, the comparisons have been based on empirical productmoment correlations, sire rankings, and the realized genetic response obtained from truncation selection. Meijering and Gianola (1985) showed that with a balanced random one-way classification and binary responses, the application of quasi-best linear unbiased prediction (QBLUP) and maximum a posteriori estimation (MAPE) leads to an identical ranking. With 4 categories of response and constant progeny group size, QBLUP and MAPE gave very similar sire rankings; the differences in the mean true breeding values of sires with either QBLUP or MAPE as selection rules were not significant. In simulation studies where the data were generated by a one-way sire model with variable progeny group size, there were noticeable differences in efficiency between QBLUP and MAPE only when the incidence was high (> 90 %, binary response) and with high (> 0.20) heritability (Meijering and Gianola, 1985;H6schele, 1988). When applying mixed models for data simulation, MAPE was generally found to be superior in comparisons with QBLUP. This superiority depends on several factors: differences in incidence, intraclass correlation, unbalancedness of the layout, and differences in the fixed effects and the proportion of selected candidates (Meijering and Gianola, 1985;H6schele, 1988;Mayer, 1991). The relationship between the relative superiority of MAPE and the proportion of selected candidates was not found to be of a simple kind (Mayer, 1991). Somewhat in contrast to these findings in the simulation studies were the results of investigations where product-moment correlations between MAPE and QBLUP estimators were calculated (Jensen, 1986;Djemali et al, 1987;H6schele, 1988;Weller et al, 1988;Renand et al, 1990). The correlation coefficients throughout were very high. This is even more astonishing because theoretically there is no linear relationship between MAPE and QBLUP, and so even with identical sire ranking the correlation coefficient is smaller than one.
Generally, a sire model can be considered as a special case of an animal model, making very restrictive supositions. For continuously distributed traits, the use of an animal model is currently the state of the art for genetic evaluation of breeding animals. Basically, all the practical and theoretical reasons in favour of the use of an animal model instead of a sire model for traits showing a continuous distribution are also valid for threshold traits.
However, in the present paper it was clearly shown that with threshold traits, the maximum a posteriori estimators of the parameters of the underlying variable or the estimators of the breeding value are different whether a sire model or an equivalent animal model is used to describe the underlying variable. This is, of course, an unfavourable behaviour of the estimation method, but as long as the ranking of breeding animals is not affected, it would not be so problematic from a breeding point of view. As mentioned above, QBLUP and MAPE yield exactly the same ranking of sires for a one-way random model with binary responses and constant progeny group size although there is no strict linear relationship between QBLUP and MAPE. However, as was shown in this study, with maximum a posteriori estimation, even the ranking of breeding animals may depend on the model type chosen.
The result that the use of equivalent sire and animal models for the underlying variable and the use of the same estimation method lead to different estimators may be counterintuitive. Under the animal model the maximum a posteriori estimators are obtained by computing the mode of the joint posterior distribution of the 'environmental' effects and the breeding values of all the animals. Under the sire model the mode of the marginal posterior density for the sires is computed. In the analyses of the simple example data set in this study, the 'marginal' posterior density for the sires equals the joint posterior density, after integrating out the genetic effects of the progenies (v l , v JJ g(t, Vl , V2 , V3 , !4 !! h 2 )8v 1 8v 2' It is well known that the posterior densities in this study are only asymptotically normal and generally do not have typical forms of distribution (see, for example, Foulley et al, 1990). Therefore the mode of the marginal posterior density function is different from the mode of the joint posterior distribution. The situation is different in the linear model case or more precisely when the likelihood follows a normal distribution. As shown by several authors (eg, Gianola et al, 1990) then, under the assumption that the variance components are known, the joint posterior density is multivariate normal, ie modes and means are identical and marginal means can be derived directly from the vector of joint means.
If the maximum a posteriori estimates depend on the model type used to describe the underlying variable, as shown in this paper, there may arise the question as to which is the better estimator. In animal breeding, the best selection rule, in the sense of maximizing the expected value of the mean of the breeding values of a fixed number of individuals selected from a fixed number of candidates, is the mean of the marginal posterior distribution of the breeding values (Goffinet, 1983;Fernando and Gianola, 1986). Only because its calculation is usually thought to be unfeasible for threshold traits (Foulley et al, 1990;Knuiman and Laird, 1990), are joint posterior modes or maximum a posteriori estimators considered and regarded as an approximation to the posterior mean. For convenience and simplicity, let us assume that in the analysis of the simple example data set in this study, the value of the threshold is known to be t = 0. The marginal posterior mean can then be calculated, not only by numerical means, but also analytically. Under the conditions that the breeding values and the environmental effects are normally, identically and independently distributed with mean 0 and variances Q a and uf respectively, and mutually independent, it follows from well-known statistical properties of the normal distribution that the marginal expected value E[alM o] is: Using the additive-genetic transmission model, where p is a progeny of sire s and dam d, and a m is a random variable with mean 0 and variance u'12 describing the Mendelian sampling: the marginal breeding value of sire 1 ( U1 ), conditional to his progeny exhibiting the realization Mo(n2!) is: In table IV this marginal mean estimator is compared with the maximum a posteriori estimators under the 2 different model types (sire model and animal model).
In the simple example data set of this study the maximum a posteriori estimator under the sire model is much closer to the marginal mean estimator. This is especially true for heritabilities greater than 0.25. With heritabilities greater than 0.50 the maximum a posteriori estimator under the animal model proved to be very poor when compared with the marginal mean estimator. In a more thorough study of marginal mean estimation by Mayer (1994b), similar results were found in the analysis of the calving ease data set presented in table I. Thus, a normal density approximation may not be justified for posterior distributions when n ij values are small as under an animal model. Further Mayer (1994b) showed that the standard deviation of the marginal posterior density was clearly greater than the approximate standard deviation of the maximum a posteriori estimates derived from the observed or estimated Fisher information matrix.