Use of covariances between predicted breeding values to assess the genetic correlation between expressions of a trait in 2 environments

Summary - Procedures to interpret correlation and regression coefficients involving predicted breeding values (BV) calculated for the same animals in different environments have been developed. Observed correlations are a function of the additive genetic correlation between performances in the 2 environments but are also affected by selection of animals that produce data in both environments, the accuracy of BV predictions in each environment, relationships among animals within and across environments and covariances among BV predictions within an environment arising from estimation of fixed effects in best linear unbiased prediction (BLUP) of animal BV. Methods to account for effects of selection and variable accuracy and experimental designs to minimize effects of relationships and covariances among BV predictions from estimation of fixed effects have been described. The regression of predicted BV in environment 2 on predicted BV in environment 1 is generally not affected by selection in environment 1, but both correlation and regression coefficients are sensitive to covariances among breeding value predictions within environments. In general, caution must be exercised in interpreting observed associations between predicted breeding values in different environments.

Summary -Procedures to interpret correlation and regression coefficients involving predicted breeding values (BV) calculated for the same animals in different environments have been developed. Observed correlations are a function of the additive genetic correlation between performances in the 2 environments but are also affected by selection of animals that produce data in both environments, the accuracy of BV predictions in each environment, relationships among animals within and across environments and covariances among BV predictions within an environment arising from estimation of fixed effects in best linear unbiased prediction (BLUP) of animal BV. Methods to account for effects of selection and variable accuracy and experimental designs to minimize effects of relationships and covariances among BV predictions from estimation of fixed effects have been described. The regression of predicted BV in environment 2 on predicted BV in environment 1 is generally not affected by selection in environment 1, but both correlation and regression coefficients are sensitive to covariances among breeding value predictions within environments. In general, caution must be exercised in interpreting observed associations between predicted breeding values in different environments. predicted breeding value / genetic correlation / regression / selection index / best linear unbiased prediction Résumé -Utilisation des covariances entre les valeurs génétiques prédites pour estimer la corrélation génétique entre les expressions d'un caractère dans 2 milieux.
Des procédures sont établies pour interpréter les coefficients de corrélation et de répression impliquant des valeurs génétiques prédites (VG) calculées pour les mêmes animaux dans différents milieux. Les corrélations observées sont fonction de la corrélation génétique entre les performances dans les 2 milieux, mais elles dépendent aussi de la sélection des animaux sur lesquels des données sont recueillies dans les 2 milieux, de la précision des prédictions de VG dans chaque milieu, des parentés entre animaux intra-milieu et entre milieux et des covariances entre les prédictions de VG dans un milieu qui résultent de INTRODUCTION Procedures to estimate additive genetic correlation (r G ) between expressions of the same trait in different environments were introduced by Falconer (1952), Robertson (1959), Dickerson (1962) and Yamada (1962). The procedures are analogoes to those for estimation of genetic correlation between 2 traits in the same environment, but recognize that performance is normally not measured on the same animal in multiple environments. Instead, related animals (often half-sibs) are produced in each environment and r G is derived by comparing the resemblance among relatives in different environments to that observed among relatives in the same environment.
In single-generation experiments utilizing half-sibs, sires can produce progeny in pairs of environments. If sires are evaluated in environment 1 before being used in environment 2, divergent selection of sires can increase precision of estimates of the genetic regression of one trait on the other when a fixed number of progeny is measured (Hill, 1970;Hill and Thompson, 1977). This strategy makes use of the fact that 'selection of sires biases correlation between parent predicted breeding value (BV) and offspring performance but does not affect the regression of progeny performance on parent predicted BV so long as there is no selection of progeny records.
Data from industry performance-recording programs often include records of relatives evaluated in different environments, but the data structure is not under experimental control. Animals differ in the amount of information available, and unknown non-genetic sources of resemblance among relatives can exist. Likewise, little information may exist on procedures used to select parents in each environment. Procedures to estimate additive genetic covariances from these industry data sets exist (Meyer, 1991) and have been used to estimate covariances between expressions of the same trait in different environments (eg, Dijkstra et al, 1990). These analyses require that the model includes all genetic and nongenetic sources of resemblance among relatives and that information used to select parents be included in the data. Large numbers of records often exist, but only a fraction of them mey represent records of relatives in different environments. Restriction of data to only records of animals with close pedigree ties across environments is tempting to reduce computational requirements, but may violate assumptions regarding selection. If predicted BV for the same animals in different environments were derived using only data from within each environment, correlations among predicted BV across environments should provide information about r c . Observed correlations in such situations (Oldenbroek and Meijering, 1986;DeNise and Ray, 1987;Tilsch et al, 1989a,b;Mahrt et al, 1990) were usually < 1, but, as noted by Calo et al (1973) and Blanchard et al (1983), the expected value of the correlations is also < 1, even if the underlying genetic correlation is unity. Thus correlations between predicted BV in different environments must be interpreted relative to their expected value. This paper will consider the expected values of observed correlations and consider alternative experimental designs. Expected values of correlation and regression coefficients under ideal conditions will first be reviewed. Effects of non-random selection, variation in the accuracy of BV predictions, relationships among animals within and across environments, and covariances among BV predictions arising from estimation of fixed effects under best linear unbiased prediction (BLUP) will then each be considered.

RELATIONSHIP BETWEEN PREDICTED BV IN 2 ENVIRONMENTS
Let a population in environment 1 have additive genetic variance Q u l for some trait. Predict BV (Mi) in that environment, and choose m sires to produce progeny in environment 2. Predict BV in environment 2 (Û 2 ) using only data from that environment. Let the additive genetic variance for the trait in environment 2 be ol u 2 2and the genetic correlation between BV in environment 1 (u l ) and 2 (u 2 ) be r G .
Let the accuracy of BV prediction, a l and a 2 for environments 1 and 2, respectively, be the correlation between actual and predicted BV and be constant within each environment.
Under certain conditions, the expected correlation between predicted BV in the 2 environments (Tulu2! is a,r G a 2 and r G can be estimated as r G = r&mdash; ! /a l a 2ul!!2) Ul 2 The conditions include (Taylor, 1983): 1) no environmental correlation between performance in the different environments; 2) no relationships among parents of measured animals; and 3) no other covariances among predicted BV within either environment. An additional assumption (4) is that sires are chosen at random.
For sire evaluation with these assumptions, a ij 2 = n ij/ (n ij + A) where n ij is the number of progeny for sire i in environment j and A is the ratio of residual to sire variance. Assumption (1) is normally met if different animals are measured in different environments. Assumptions (2) and (4) can be met through choice of sires. Assumption (3) will not normally hold for BLUP, but may approximately hold under some conditions. The regression (b) of u 2 on ill has the expectation: such that TO = b!2u1 (Q!1 /a2!!z ). Taylor's (1983) assumptions are required for this expectation, but random selection of sires is not. Knowledge of a!l and Q u 2 is required to calculate the expected regression coefficient and for prediction of u i and u 2 . Effects of using incorrect values of o,2 on estimates of r G will not be considered further, but may be important. Expected confidence limits for observed correlation (F) and regression coefficients (b) can be used to evaluate experimental designs in terms of their ability to detect significant departures of r and from their expectations. For correlation analysis, Fischer's z = 0.5[ln (1 +r) -In (1 -r)] (Snedecor and Cochran, 1967) has variance of ! (m-3)-1 where m is the number of sires in the sample. For ) r) 6 0.65, confidence bounds on F are of similar width at fixed m, whereas for Irl > 0.65, confidence bounds narrow with increasing r. Large numbers of sires are thus required if accuracies are low to avoid confidence limits that overlap 0.
For regression analysis, variance of b [V(b)] is: (Snedecor and Cochran, 1967) where Q!2I!1 is the variance in Û 2 at a fixed value of ! U2 Ut Û l (ie, the mean square for deviations from regression) and SS(ul) is the sum of squares for u l . Given T_aylor's (1983) assumptions for m sires sampled at random from environment 1, V(b) is: Numbers of sires and progeny required to detect significant departures of 6 from its expected value are given in figure 1 for several values of rc, a u , , O &dquo; U2 and a l = a 2 or a l = 0.95. When a l = a 2 (eg, when sires are being proven simultaneously in 2 environments), numbers of progeny required in each environment to reject the null hypothesis that r G = 1 are minimized at a j = 0.7 to 0.8 for r G between 0.5 and 0.8. For a l = 0.95 (eg, when proven sires are chosen from environment 1), progeny numbers in environment 2 are minimized with one progeny per sire, but increase little until a 2 exceeds 0.5 to 0.6. Thus relatively efficient designs at a l = 0.95 would include 35 to 45 sires with 400-500 progeny at r G = 0.6, but 250-400 sires with 1200-1300 progeny at r G = 0.8. Critical numbers required for correlation analysis were similar to those for regression analysis at low accuracies and r G , but lower at higher accuracies due to asymptotic declines in the width of confidence intervals as expected r increased. The ratio of the critical number of sires for correlation analysis to critical number for regression analysis (SRAT) was predictable (R 2 = 0.983) as a function of q = a l a 2 and r e such that SRAT = 1.115-0.101 q-0.667 q 2 -0.161 1 rG . This ratio adjustment can be applied directly to values in figure 1 to approximate critical sire and progeny numbers for correlation analysis.
The above derivations assume that accuracies are calculated correctly in both environments. Under BLUP, accuracies of u i for non-inbred animals are given by (1 &mdash; Ci2/Q!)'S where C ii is the ith diagonal element of C 22 , the prediction error covariance matrix of u (Henderson, 1973). If the model is complete and properly parameterized, accuracies are expected to equal correlations between actual and predicted BV. In most applications, u is derived by iterative solution of Henderson's (1963) mixed model equations (MME) rather than by direct inversion. Diagonal elements of C 22 are approximated but off-diagonal elements of C 22 are usually not estimated. To date, no completely satisfactory procedures to obtain diagonal elements of C 22 exist. Alternative methods have been presented by Van Raden and Freeman (1985), Greenhalgh et al (1986), Robinson and Jones (1987), Meyer (1989) and Van Raden and Wiggans (1991). Evaluation of procedures to estimate accuracy is beyond the scope of this study, but the assumption that accuracies are estimated correctly is critical to the discussion.
Effects of departures from the ideal conditions described above will now be discussed.
Each scenario was repeated for r G = 1.0 or 0.5 and replicated 10 times. Each replicate contained 5 000 selected animals.
Agreement between predicted and simulated values of V(ii 2 ) and Corr(ic l u 2 ) (table I) was within theoretical 95% confidence limits of the expected value (Snedecor and Cochran, 1967). Thus equation [3] predicted Corr(û l û 2 ) satisfactorily in bulls selected non-randomly on u l with fixed accuracy a l .
With selected sires, the expected V(b) is: using values from equations [1] and [3]. The SD of is inversely proportional to vi --w and varies from 48 to 243% of its value when w = 0 as w varies from -3.39 (divergent selection from the top and bottom 5% of the population) to 0.83 (selection from the top 10% of the population). Sample sizes to detect significant departures of Regr(u 2 u1) from its expectation using selected sires can be derived from figure 1 by dividing sire and total animal numbers by 1w.
Effects of variation among animals in accuracy of predicted BV Calo et at (1973) and Blanchard et at (1983) derived the expected correlation between predicted BV for 2 traits when BV for each trait were estimated in separate single-trait analyses and individuals differed in accuracy of BV prediction as C ' OT'r(ui M2 ) = r G a l2 for: (see Appendix) and recommended using this expression to estimate r G from observed C O rr(iil i i 2 ). Similarly: Taylor (1983) criticized equation [5] as unstable, however, asserting that it may yield estimates of r G that are outside the parameter space, and presented assumptions required to allow estimation of r G with this equation. Taylor (1983) concluded that, if all assumptions are met, equation [5] is appropriate to estimate r G so long as the a! are derived from MME as (1 -C,,/ 0 ,2). The equations of Calo et al (1973) and Blanchard et al (1983) do not consider selection on Ûl. With selection, equation [5] appears appropriate to estimate Corr(u l u 2 ) for the selected sample, but not within the unselected population. Equations [3] and [5] could, however, perhaps be combined to give the expected correlation in a selected sample of animals with variable accuracy as: To evaluate equation [7], several accuracy scenarios (AS) were considered by simulating samples from the u i distribution. AS1: 20 000 animals from the upper 10% of the u l distribution. a l varied uniformly over the interval 0.7 to 0.95. AS2: 20 000 animals from the upper and lower 10% of the ui distribution. a l varied uniformly over the interval 0.7 to 0.95. AS3: 10 000 animals from the upper 10% of the Û l distribution with a l uniformly distributed over the interval 0.7 to 0.95 and 10 000 animals selected from the lower 10% of the distribution with a l uniformly distributed over the interval 0.5 to 0.7. AS4: 10 000 animals from the upper 10% of the ui distribution with a l uniformly distributed over the interval 0.7 to 0.95 and 10 000 animals selected from the bottom 80% of the distribution with a l uniformly distributed over the interval 0.5 to 0.7. ASS: 15 000 animals from the upper 10% and 5 000 animals from the lower 80% of the Û l distribution with a l uniformly distributed over the interval 0.7 to 0.95. AS6: 5 000 animals from the upper 10% and 15 000 animals from the bottom 80% of the Û l distribution with a l uniformly distributed over the interval 0.5 to 0.995. AS7: 10 000 animals from the upper 10% of the Û l distribution with a l uniformly distributed over the interval 0.795 to 0.995 and 10 000 animals from the lower 80% of the distribution with a l uniformly distributed over the interval 0 to 0.50. a 2 was uniformly distributed over the interval 0.5 to 0.7 for all scenarios. Each scenario was repeated for r c = 0.5 or 1.0 and replicated (table II). Empirical calculation of w requires use of a2 , which varies with accuracy. Simulated values ! u l of Mi were thus standardized by dividing by a li a ul and empirical w calculated as 1 &mdash; V(u l ) using standardized u l .
For all accuracy scenarios, observed Regr(u 2 u1) agreed closely with predicted values from equation [6] (table II). Observed values of Corr(u l u 2 ) were usually also close to expectations from equation [7], but with some systematic departures from expectations. For directional selection (ASl), the mean observed Corr(u l u 2 ) was slightly but significantly larger than predicted (by 0.010 t 0.002 for both r G ).
Thus equation [7] produced a small negative bias under directional selection with variable accuracy. This result was confirmed by producing 10 more replicates at r G = 1; the results were identical.
For divergent selection, differences between observed and predicted correlations were again small, but sometimes significant and now negative for AS2, 3, 5 and 6, ranging from -0.001 to -0.008 (!0.002). However, with both non-symmetrical selection from high and low groups and different accuracy distributions between groups (AS4 and AS7), observed correlations were considerably larger than predicted, especially for r G = 1 (table II). The appendix shows exact expectations for correlations and regressions involving u i and Û 2 under non-random selection from environment 1 and variable accuracies within environments. Correlation between means and accuracies of divergently selected groups violate some of the assumptions used to derive equation [7] and presumably account for the departures from predicted values in AS4 and AS7. Equation [7] thus produced slightly biased predictors of Corr(Ei E2 ) but still appears useful, especially when exact selection rules are unknown. However, biases in predicted values of Corr(û l Û2 ) in equation [7] will be multiplied by the inverse of the coefficient of r G in equation [7] to estimate r G . Potential bias in re thus is larger with lower a j or more directional selection. If V(u i ) is larger than expected from random selection and greater precision than that provided by equation [7] is desired, Ap P endix equations can be used.
The expected correlation between Û l and Û 2 thus depends on the distribution of accuracies within each environment, the selection applied on Û l (quantified by w). and r G . To evaluate net effects of these variables, values of Corr(û l û 2 ) from equation [7] were calculated for r G = 1 and when a 12 varied from 0.10 to 0.90 and w varied from -3.3 to 0.9 (table III). For r e = 1.0 and a 12 = 0.90, Corr(u l u 2 ) varied from 0.55 to 0.97 due to selection, although r G exceeded 0.79 so long as directional selection was not intense (w 6 0.6): For w = 0 and r G = 1.0, Corr(u lu2 ) equalled a l2 , but still varied from 0.10 to 0.90, depending on observed values of a l and a 2 .

Effects of relationships
Previous results assume independence of predicted BV within each environment. However, relationships among animals lead to covariances among predicted BV within and across environments. If ui and U 2 are predicted by BLUP, covariances among predicted BV within environments also arise from estimation of fixed effects. These covariances affect expectations of both Corr(u l u 2 ) and Regr(u 2 u1).
Their impact is difficult to generalize, depending upon the extent and nature of relationships in the data and the distribution of records among fixed effect classes. Covariances among predicted BV associated with relationships and estimation of fixed effects arise simultaneous in BLUP solutions to MME, but effects of relationships alone can be seen under selection index, or best linear prediction (BLP), assumptions of known mean and variance for both u l and u 2 . In that case: where Û j is a vector of breeding value predictions for environment j and y j is the data vector in environment j with covariance matrix V j . H j is the covariance matrix between y j and u'.. For non-inbred animals, H j = Z j G = Z j Ao, 2 where Z j is the incidence matrix relating y j to Uj , and G and A are additive covariance and numerator relationship matrices, respectively, for animals in Uj . The covariance matrix of Gj (Qj ) and the covariance matrix between Û l and u2 (Q i2 ) are thus: Expectations of sample variances of U j (s3-) and covariance between M i and _ Uj u2(Sulu2) are functions of elements of Q j and Q 12 : where tr is the trace of the matrix and sum is them sum of all elements. If animals recorded in the 2 environments resemble one another only because of relationships to animals sampled from environment 1, This assumption is warranted if animals are evaluated using unrelated populations of mates in the 2 environments but may not be correct if mates are potentially related across environments. If selected animals are likewise unrelated, and relationships among recorded animals within each environment arise only through relationships to selected animals, Q j will be diagonal with elements a? . 2 for the ith animal in the jth environment, A is an identity matrix of size m and Q lz is diagonal with elements aiiaiirGaul!u2. Sample correlation and regression coefficients then have expectations equal to those previously discussed.
In most applications, Q j are not diagonal and off-diagonal elements are not calculated due to the nature and size of V! 1. Thus explicit consideration of offdiagonal elements of Q j may not be possible unless the data set is small or highly structured. However, if the accuracy of all Û ij approaches 1, Q j approaches Ao, &dquo;j 2 (ie, Qu --! or2 u j ) and Q 12 approaches A rcO &dquo; utO &dquo; u2 such that Corr(û l û 2 ) = TG and°/ J ' Regr(û 2 û¡) = r C O&dquo;U 2/ O &dquo; U t' A more realistic situation is one in which sires from environment 1 have a, approaching 1 but are evaluated in environment 2 with a 2 < 1. In this case, Q l = AQ!I,Qz = H2VZ l Hz, and Q 12 = Q 2 (rc O &dquo; Ut/O &dquo; U2 ) ' These quantities, and associated expected sample variances and covariances, could be obtained if the size or structure of the data allows calculation of all elements of Q 2 and would allow derivation of an exact predicted value for C'orr(uiU2)-A small example will demonstrate the impact of relationships on Corr(ûlû2) and Regr(û 2 ûl), which are equal in these examples. Let h z = 0.25, Qu , = O &dquo;U 2 = 1 and r G = 1. Let 3 sires produce 8 progeny each in each of 2 environments. If sires are unrelated and progeny are related only through the sires, sample variances and covariances involving Û l and Û 2 equal expected population values, and Corr(û l Û 2 ) = a l a z . If all 3 sires are full sibs, Q j and Q lz are no longer diagonal and a l = a 2 = 0.643. The expected Corr(ûlû2) is 0.211 versus a l a 2 = 0.414. If sires are half sibs, Corr(û l Û 2 ) = 0.286 versus a l a 2 = 0.365, reducing bias by one half as relationships among sires decline. Still, with many close relationships among sires, a l a 2 may considerably overestimate the expected correlation.
If only 2 of the 3 sires are full sibs, bias is reduced. a l = a 2 = 0.621 for related sires and 0.590 for the unrelated sire, and a 12 = 0.374 (equation [5]). Now a 12 overestimates the observed correlation by only 3.9% (0.374 versus 0.360). Thus if sampled animals represent a reasonable number of unrelated families, Q j and Q 12 are correspondingly sparse and little bias in Corr(u i U 2 ) is expected.
Turning to effects of relationships within environments, let the 3 sires be unrelated but cross-classified within each environment with only 2 dams. In that case, a l =az =0.566 and Corr(ii IU2 ) = 0.364 which is 13% larger than a l a 2 = 0.321.
However, under the more realistic assumption of cross-classification with 2 maternal grandsires, Corr(ûl Û2) = 0.352 versus a l a z = 0.338, yielding little bias. When sires were cross-classified with 8 dams, Corr(û l û 2 ) = 0.364 versus a l a 2 = 0.349, again yielding little bias. These examples suggest selection of widely proven, lowly related animals from environment 1 followed by evaluation in environment 2 using a broad sample of mates.

Effects of covariances arising from estimation of fixed effects
If fixed effects are estimated simultaneously with BLUP of iij, Mallinckrodt (1990) noted that off-diagonal elements of Q j and Q 12 are not zero, even in the absence of relationships. By BLUP: f or # = (X!V! 1X!)-1X!V! ly! and P_, = X!(X!V! 1X!)-1X! and wherep is the vector of fixed effects with incidence matrix X.
The covariance matrix of Û j is: composed of a term due to BLP relationships minus a term due to estimation of fixed effects. If relationships across environments arise only through sires sampled from environment 1 such that Cov(y l y2) is given by equation (10], Q 12 is still given by equation [11], but regardless of the relationship structure of the data, Q j now approaches a diagonal matrix only as a j approach 1. Also, for a given number of progeny, a ij will be less for BLUP than for BLP and depends on the number of sires and their distribution among fixed effect classes. The above solutions are identical to those obtained from MME (Henderson, 1963(Henderson, , 1984 such that: The impact of fixed effect estimation on Corr(û l û 2 ) can be seen most readily using MME for a sire model without relationships among animals in u and where P includes only contemporary group effects. Note that in all remaining examples, Corr(£i E2 ) = Regr(u 2 u 1 ) when a l = a 2 . For such a model, after absorption of fixed effects into u equations and factoring of residual variance from both sides of the equation, the coefficient matrix for u has: (Do, 1991) where A is the ratio of residual to sire variance and n i .,n. k and n ik are numbers of records for sire i, contemporary group k (of g) and sire x contemporary group subclass ik, respectively. For balanced data, n ik = n for all i and k, diagonals reduce to [gn(m -1)/m + A][vs(gn + A) for BLP] and off-diagonals reduce to (-gn/m) [vs 0 for BLP]. The corresponding inverse of the coefficient matrix has diagonal elements of (gn + mA)/[m(gn + A)] and offdiagonals of gn/[mA(gn + A)] (Searle, 1966). Q will have corresponding diagonal and off-diagonal elements of gn(m &mdash; 1)/!m(gn + A)] = a 2 and -gn/[m(gn + A)], respectively. s 2 is thus gn/(gn + A) = a !(ml)/ 7 7t]. If design matrices are the same for both environments and r G = 1, Q 12 will have diagonal elements of 9 2 n 2 ( M _ 1)/[,rn(gn+A) 2 1 = a 4 m/ (m-1) and off-diagonals of -g 2 n 2 /[m(gn+À? ] = -a9m/(m -1) to give S12 = g2 n 2/(g n + A) 2 = a 4 m 2 /(m -1) 2 and Corr(ûlû2) = gn/(gn + A) = a 2 m/(m -1). With balanced designs, Corr(u l u 2 ) has the same expectation under both BLUP and BLP, but accuracies are lower under BLUP such that a l a 2 from BLUP underestimates expected Corr(û l û 2 ). The extent of bias is proportional to m/(m&mdash;1) and decreases from 20% at m = 5 to 11% at m = 10 and 2.6% at m = 40. This expectation is maintained if design matrices differ between environments provided designs are balanced within each environment. The situation is more complicated for unbalanced designs, but general conclusions are similar in that the number of sires compared as contemporaries needs to be large enough to minimize confounding between sire BV predictions and fixed effects estimates. Otherwise, BLUP accuracies are reduced and their product underestimates expected Corr(u l u 2 ). For example, consider a block of 8 sires with progeny distributed over 4 contemporary groups (eg, 4 yr of an experimental evaluation in some environment) as shown in table IV. The size of the experiment may be varied by increasing the number of sire blocks (to 16, 24, etc, sires), by varying the number of progeny per sire and contemporary group (n), by replicating the sire block over additional contemporary groups (8, 12, 16, etc), or by a combination of these approaches. The same design is assumed for each environment.
Define bias (fig 2) as the difference between the expected Corr(u l u 2 ) calculated from equations [8] and [9] and the product a l a 2 which is constant for all sires in this design. With n = 6 and only 8 sires, bias was relatively large with Corr(£1 £2 ) = 0.41 vs a l a 2 = 0.35. Bias decreased as number of sire blocks increased and was < 0.03 with 24 sires (12/contemporary group). Doubling n or replicating sire blocks across more contemporary groups (not shown) did little to change the pattern of bias. Expected values of Corr(ûlû2) were also compared to the product of BLP accuracies calculated ignoring contemporary group effects. The product of BLP accuracies overestimated Corr(u l u 2 ) as shown by negative bias in figure 2, but the product of BLP accuracies was superior to the product of BLUP accuracies as an estimator of Corr(u l u 2 ).
Design like that in table IV can be used under experimental conditions, but are less feasible when sires are compared on cooperator farms. It particular, use of large numbers of experimental sires on individual farms may not be feasible. Instead, sires from environment 1 may be tested together on several farms in a loosely connected design but with only a few sires represented on any one farm. If only data from introduced sires are used in the evaluation, considerable bias in Corr(u l u 2 ) may result. However, this bias can be reduced if introduced sires are evaluated with sires represented only in environment 2 and data from all sires are included in the evaluation. To demonstrate this effect, two additional sires were added to each contemporary group in table IV to give 16 sires/sire block with n = 6. Added sires produced progeny in only one contemporary group. Bias in resulting values of Corr(ul u2 ) for introduced sires was reduced (fig 2). Thus if sires from one environment are introduced into another and if evaluation occurs primarily in cooperator herds, sires should be evaluated in contemporary groups containing reasonably large numbers of sires (either introduced or native) to increase precision of estimates of contemporary group effects, and data from all sires should be included in the evaluation.

CONCLUSIONS
Interpretation of correlations between predicted BV in different environments is not straightforward. Expectations of such correlations are influenced by accuracy of evaluation of animals in both environments, by selection of animals chosen for evaluation, by relationships among chosen animals and by the design of the evaluation in both environments. If animals are chosen from environment 1 for evaluation in environment 2, Regr(£2£1 ) may be a more useful statistic than Corr(u l u 2 ) because it is unbiased by selection on iij. However, both Regr(û 2 û¡) and Corr(£ 1£ 2 ) are biased by covariances among predicted BV within environments. Also, use of proven sires (a -! 1.0) from environment 1 simplifies interpretations and reduces the number of sires required to attain a specific level of significance for measures of association.
Equations in this paper allow calculation or approximation of expected values of Corr(£ 1 £ 2 ) under various sorts of selection and with variable accuracies in each environment (equation [7]). Evidence for selection can be obtained empirically if necessary by comparing observed V(Û¡) to its expectation from Blanchard et al (1983) (see Appendix). Expected values of both Corr(£ 1 £ 2 ) and Regr(û 2 û¡) may involve off-diagonal elements of Q j matrices which are often not available for BLUP BV predictions. Effects of off-diagonals may be minimized by ensuring that a number of families are represented in both the experimental animals and their mates and by using several (eg 8-16) sires per contemporary group. Use of small numbers of sires per contemporary group can lead to considerable underestimation of the expected value of Corr(£1 £2 ) if off-diagonal elements are not considered. Prediction of expected values of Corr(£ 1 £2 ) and Re g r(ii 2 iii) from observed accuracies may be superior when selection index (BLP) rather than BLUP accuracies are used.
A number of potential difficulties in deriving and interpreting Corr(û l Û 2 ) have not been explicitly considered. Accuracies are assumed to be properly calculated, even though approximations are normally used and are probably not completely satisfactory. Additive genetic variances must be known for both environments in order to calculate u l and U 2 correctly and to interpret Regr(u 2 u 1 ). If sires introduced into environment 2 for evaluation are a selected sample from environment 1 and BLUP evaluations are used, grouping strategies and(or) adjustment of covariances may be required to derive unbiased ii 2 -Values of o,2 U2 for animals selected from environment 1 will depend both on selection applied and on r G . Thus results given for correlations involving BLUP predictions of Û j are probably striclty correct only for random selection of sires. See Diaz (1992) for additional discussion of effects of selection and grouping on Corr(£ 1 £ 2 ) with BLUP predictions. Despite these problems, however, correlations and regressions involving BV predictions in different environments will often be relatively easy to obtain and, if properly interpreted, can provide information on r G .

APPENDIX
This appendix addresses general expectations for variances and covariances of ill and M 2 with non-random selection on Û l and variable accuracies in both environments. Sample m animals from environment 1. If sampling is non-random and selection rules are not specified or if accuracies differ, the resulting distributions of u j and M z are mixtures of m unique distributions. The sample sum of squares (SS) of u i is: The expected value (E) and variance of Û 2i with accuracy a 2i are: The resulting expected SS(u 2 ) and sum of cross-products (SCP) are: Expected correlation and regression coefficients can be derived from these SS and SCP. If animals are chosen at random, u l and U 2 have mean 0 and expected Corr(£ 1£2 ) is given by text equation [5] (Blanchard et al, 1983).
If selection is non-random but accuracies are constant within environments, formulae for SS(u 2 ) and SCP(u l u 2 ) become consistent with text formulae 2 and 3. For directional selection with constant accuracies, w can be replaced by i(i &mdash; x) without loss of generality. For divergent selection, let fraction w H of the selected animals come at random from the upper fraction O H of the distribution with standardized mean and truncation point of i H and x H , respectively. Let fraction W L come from the lower fraction ø L . Let accuracies within high and low groups, and within environment 2, be constant at a lH , a lL and a 2 , respectively. V(û¡) within selected groups are [1i H (i H -x H )]aî H O&dquo;î and [1i!(iL -x L )]aî L O&dquo;I. The expected value and variance of u l and u 2 , and their expected covariance are: The resulting Corr(û l û 2 ) reduces to text equation [4] only if a lH = a lL and selection is symmetrical (ie, w H = WL and i H = -i L ).