Search
Author
Title
Vol.
Issue
Year
1st Page

Journal of Animal Science - Animal Genetics

Social genetic effects influence reproductive performance of group-housed sows1

 

This article in JAS

  1. Vol. 93 No. 8, p. 3783-3793
     
    Received: Mar 19, 2015
    Accepted: June 01, 2015
    Published: July 24, 2015


    2 Corresponding author(s): kbunter2@une.edu.au
 View
 Download
 Share

doi:10.2527/jas.2015-9111
  1. K. L. Bunter 2*,
  2. C. R. G. Lewis and
  3. S. Newman
  1. * Animal Genetics and Breeding Unit (AGBU)3, University of New England, Armidale, NSW 2350, Australia
     PIC Europe, Genus PIC, PIC Espana, Avda. Ragull, 80, 08173 Sant Cugat del Valles, Barcelona, Spain
     Genus Plc, Hendersonville, TN 37075

Abstract

Group housing of gestating sows has implications for reproductive performance due to detrimental interactions between sows within groups. Reproductive records (n = 10,748) were obtained for 8,444 pedigreed nucleus sows housed in a single facility, formed into 1,827 static groups during gestation. Only data from complete groups were used to estimate genetic parameters for total born (TB), number born alive (NBA), and gestation length (GL) and to compare models extended to account for group effects. Censored data for sows which did not farrow (0.8% of records) were augmented with biologically meaningful values. Group sizes ranged from 2 to 10, in pens designed to hold 4, 8, or 10 sows per pen. Sows were grouped by parity, line, and mating date after d 35 of pregnancy. Heritability estimates were generally constant across all model alternatives at 0.11 ± 0.02 for TB and NBA and 0.32 ± 0.03 for GL. However, models for all traits were significantly (P < 0.05) improved through inclusion of terms for nongenetic group and social genetic effects (SGE). Group effects were no longer significant in models containing both terms. The proportional contributions of SGE (s2) to phenotypic variances were very low (≤0.002 across traits), but their contributions to calculated total genetic variance (T2) were significant. The differences between h2 and T2 ranged between 3 and 5% under simple models, increasing to 8 to 14% in models accounting for both covariances between additive direct (A) and SGE and the effects of varying group size on the magnitude of estimates for SGE. Estimates of covariance between A and SGE were sensitive to the modeling of dilution factors for group size. The models of best fit for litter size traits used a customized dilution based on sows/pen relative to the maximum sows/pen. The best model supported a reduction in SGE with increased space per sow, independent of maximum group size, and no significant correlation between A and SGE. The latter is expected if A and SGE reflect different trait complexes. It is suggested that the SGE estimated for reproductive traits represented the expression of an unobserved phenotype, such as sow aggression, of an individual on its pen mates. Further investigation into the use of competitive effects models for genetic evaluation of reproductive traits for group-housed sows could be considered a strategy to improve welfare and performance of group-housed sows.



INTRODUCTION

Group housing of gestating sows has become common in Australia. Group housing improves some aspects of sow welfare over individual housing by enabling exercise and social interactions between sows, thereby improving cardiac health, leg strength, and reducing stereotypic behaviors (Marchant and Broom, 1994; Barnett et al., 2001; Karlen et al., 2007). However, group housing also enables detrimental social interactions, including acute aggression at mixing and feeding and chronic stressors (Barnett et al., 2001). Aggressive encounters can result in decreased welfare and performance due to lameness and injuries, poor body condition, and embryonic loss (Brown and Seddon, 2014), and ongoing stress can also affect immune response measures (Morrow-Tesch et al., 1994; Karlen et al., 2007). In addition to using appropriate housing design, group formation, feeding strategies, and management to minimize aggression between sows (Brouns and Edwards, 1994; Marchant-Forde and Marchant-Forde, 2005; Li et al., 2012; Camerlink et al., 2013; Hemsworth et al., 2013), it is possible to consider that some sows are better adapted to group housing and selection can also occur at this level (Newman, 1994). Social effects that are inherited are termed social genetic effects (SGE; Muir, 2005; Bijma, 2010b), and significant estimates have been obtained for laying traits in poultry (Muir, 1996a,b) and the growth of Japanese quail (Muir, 2005) and pigs (Bergsma et al., 2008). Until recently (Bunter et al., 2014), no studies investigated the consequences of social interaction within groups for reproductive traits in pigs because historical data reflected performance under individual housing. Sows are potentially regrouped several times across gestation events. The aim of this study was to establish whether sow grouping during gestation had a significant impact on reproductive outcomes and whether SGE were present for total born (TB), number born alive (NBA), and gestation length (GL).


MATERIALS AND METHODS

Data

Sow Grouping And Management

These data, commencing in 2010, were from pedigreed nucleus sows in a single facility located in the northern midwest of the United States, group housed during gestation, with pen and date of grouping recorded. All sows were bred using AI and held in individual stalls until scanning to confirm pregnancy status at around 28 d after mating. Sows that returned to service or that were not pregnant at scanning were remated or culled according to farm protocols and were not allocated to a group pen for gestation until confirmed pregnant. Pregnant sows were grouped together around d 35 of gestation, predominantly on 1 date per pen only, by considering mating day, sow parity, and sow line (maternal versus terminal lines). Gestation pens were available in 3 sizes to house a maximum (nmax) of 4, 8, or 10 sows per pen. Regardless of pen size, the space allowance per sow was identical (1.49 m2/sow) when the pen was full, and all pens had shoulder barriers and separate feed delivery per sow to help prevent sow displacements during feeding events. No pen held more sows than the maximum they were designed for. However, if there were not enough sows to group together on any given day for the pens available, the number of sows housed per pen could be lower than the defined maximum group size.

Sows in their first gestation were typically grouped separately to later parity sows, and sows closest in mating dates were preferentially grouped together for efficient use of the pens available. Overall, 60% of sows were in pens representing only a single parity grouping. Maternal line sows were generally grouped separately to terminal line sows, the latter of which were also typically housed in smaller groups. Sows could occasionally be removed from a gestation pen during the gestation period, but no replacement sow would be introduced into that group. Therefore, groups were essentially static from allocation to farrowing unless a sow was removed, but the extent of such removals was not fully recorded. Gestation pens were emptied over 1 to several days as individual sows were removed from the gestation pen at around d 110 of gestation for transfer to farrowing accommodation.

Reproductive Data

Reproductive data for sows that were mated included the TB, and the component traits include NBA, still born (SB), or mummified (MUM) piglets. Gestating sows that did not farrow were also identified with an alternative outcome for each mating event. Gestation length was calculated for sows that farrowed as the interval between mating and farrowing dates. However, trait values for farrowed and unfarrowed sows are both required, and a small percentage of sows (0.8%) did not farrow after allocation to gestation pens. Sows without a farrowing record were allocated a phenotype of 0 for TB and NBA because they represent failed-to-farrow sows. Gestation length data for these sows was the interval between mating and outcome dates when the interval was ≤109 d (i.e., the shortest gestation length in the data) or set to 110 when >109 d, since pregnancy loss had occurred before the time of transfer to farrowing accommodation, which occurs at around d 110. A similar data augmentation strategy has been used previously to accommodate censoring due to reproductive failure for defining calving interval traits in cattle (Johnston and Bunter, 1996). However, data for the number of stillbirths and MUM piglets could not be sensibly augmented from a biological perspective for sows with no farrowing outcome, and these traits were subsequently not analyzed for this study.

Data Editing for Completeness of Groups

Social genetic effects are indirect effects of an individual on the phenotypes of its pen mates. Therefore, to estimate SGE correctly, only complete groups with phenotypes can be analyzed, so unnecessary editing for outliers or because of repeated records must be avoided. As an alternative, all records of the group are discarded. Based on biological norms, no reproductive traits for sows were identified as outliers for removal, and therefore only incomplete groups were identified, where possible, for removal. Information on the allocation of individual sows to gestation pens (n = 13,747 records) was merged with outcomes from each mating, including sows that ultimately failed to farrow. Dates of entry into and exit from the pens were used to confirm, where possible, the validity of the grouping. Only data from pens containing valid group sizes (from nmax − 2 to nmax) were retained for analyses. The approximate percentages of records from groups of size 2 to 3, 4, 6 to 7, 8, and 9 to 10 were 6.9, 13.2, 41.4, 38.0, and 0.5%, respectively. These groups align with the 3 pen sizes for nmax: 2, 3, 4 for nmax = 4; 6, 7, 8 for nmax = 8; and 9 and 10 for nmax = 10. After editing for group size, there were 10,748 reproductive records from both purebred lines and a small proportion of line-cross matings, representing 8,444 sows in 1,827 gestation groups. This was about 78% of the original group data presented, demonstrating the difficulty of setting up data recording systems to continuously track group allocation for breeding sows and the substantial impact of complete group removal on data volumes.

Approximately 20.8% of sows had more than 1 reproductive record in the final data. For sows retained in the data file, additive relationships were traced back for 4 generations, giving a total of 15,205 individuals. Sows with reproductive data were the daughters of 962 sires and 4,290 dams, and 1,918 sows had both records themselves and also daughters with data.

Statistical Analyses and Models

Systematic effects for all traits and all models included the year-month of mating (37 levels), line group of the sow (including piglet line cross: 10 levels), and parity group (5 levels).

Random Effect Models

Analyses of data for all sows and all traits were first conducted using models frequently employed for reproductive traits (model 1, Table 1), which excludes terms for gestation group. Random effects were fitted for each sow under an animal model to estimate additive genetic effects and to account for repeated records per sow (permanent environmental effects).


View Full Table | Close Full ViewTable 1.

A summary of alternative random effect models applied to the data

 
Model Random effects1 Description
1 A + PE + E Typical repeated records model. Ignores group structures.
2 A + PE + P + E Model 1 + nongenetic spatial or pen specific effects, estimated across groups
3 A + PE + G + E Model 1 + nongenetic group effects. Group is effectively pen ´ subset sows housed together until farrowing
4 A + PE + G + S + E Model 1 + group + social genetic effects
5 A + PE + S + E Model 4 excluding nongenetic group effects
6 A + PE + G + S + En Model 4 + heterogeneous residuals
7 A + PE + S + E Model 5 + covariance (A, S)
8 A + PE + Sd + E Model 5 + covariance (A, S) + dilution (Sd)
9 A + PE + S + En Model 5 + covariance (A, S) + heterogeneous residuals
10 A + PE + Sd + En Model 5 + covariance (A, S) + dilution (Sd) + heterogeneous residuals
1A, additive genetic; PE, sow permanent environment; P, pen; G, gestation group (in pen); S, social genetic; E, residual; Sd, social genetic effects modeled with dilution factors; En, residuals modeled to n levels, as defined by nmax.

Social genetic effects were then included in the models through fitting 2 additional random terms: 1) a nongenetic group effect and 2) an indirect SGE (model 4, Table 1). These models, previously applied to growth traits in other studies (Muir, 2005; Chen et al., 2007; Bergsma et al., 2008), contain the first term to remove nongenetic effects that typically affect growth performance of groups of animals, such as spatial shed and specific pen-related effects (location, disease exposure), for example. However, for reproductive traits, these effects are not generally expected. Therefore, the necessity of fitting the additional nongenetic group term in the model was also investigated by substituting pen identity (312 levels) with group identity (1,827 levels) to assess the significance of pen vs. group effects per se (models 2 and 3).

The full animal model containing social effects (model 4, Table 1) is therefore represented by y = Xb + Z1a + Z2pe + Z3g +Z4s + e, where y is the vector of observations; X, Z1, Z2, Z3, and Z4 are incidence matrices relating records to effects b, a, pe, g, and s; b is the vector of solutions for fixed effects; a and s are vectors of solutions for additive and SGE with variances (∼N(0, Aσ2a) and ∼N(0, Aσ2s) or with nonzero covariance Aσas; g and pe are nongenetic group and permanent environmental effects with variances ∼N(0, I3σ2g) and ∼N(0, I4σ2pe); e is the vector of residuals ∼N(0, Iσ2e); A is the matrix describing additive genetic relationships between animals; and I3 and I4 are identity matrices. The random group term replaces fitting a residual covariance within groups (see Mrode, 2014). A reduction from the full model, excluding nongenetic group effects (Table 1), was also considered.

Under the full model, the phenotypic variance was calculated post-estimation as σ2p = σ2a + (n − 1)σ2s + σ2g + σ2pe + σ2e (Bijma, 2010b) where n was assumed to be 6 (the average of group sizes in this data), and where σ2g already contains nongenetic covariances between pen mates (Mrode, 2014). The contribution of the combined additive and social effects to total genetic variance, relative to σ2p, was given by T2 = [σ2a + (n − 1)2σ2s + 2 × (n − 1)σas]/σ2p, where σas is the covariance between additive and SGE. The corresponding total genetic merit (TGM) of an individual can be calculated (after Bijma et al., 2007) as TGMi = ai + (n − 1)si, using the average group size (n), even though the individual may have been recorded in groups of size different from n. This TGM represents the additive merit of the individual for the trait analyzed (ai), along with its contribution (si) toward the phenotypes of other sows in a group for the same trait, expressed in common trait units. However, the term si can reflect expression of a different trait entirely (e.g., aggression), as described by Trubenova and Hagar (2012), and therefore the T2 parameter can reflect multiple sources of genetic contributions to expression in a single trait. Homogeneous errors and an equal effect of an individual on performance of all pen mates, with magnitude of effect independent of group size, are implicit assumptions for a basic model including SGE. Model 5 was used to test the impact of removing nongenetic group from the complete model, while model 6 was used to examine if there were heterogeneous variances for small vs. large groups (≤4 vs. 4 to 10 sows, in 2 classes) in a complete model. For models 1 to 6, the covariance between additive and social effects was not estimated and assumed to be 0.

Dilution Factors and Covariances Between Direct and Social Effects

The effect of variable group size on the variance of SGE was considered by applying dilution factors (Bijma, 2010b) to the construction of Z4. Dilution factors are used to enable estimates of social effects to vary with group size. The most parsimonious model for each trait including SGE identified from models 4 to 6 was expanded to both estimate the covariance between additive direct and SGE and to include dilution factors (models 7 to 10). Dilution factors (d) were assessed in 0.2 increments from 0 to 1 at an average group size () = 6, using the equation Z4 = [(− 1)/(n − 1)]d, where n is the number of sows in each group, and Z4 is the coefficient used in the Z4 design matrix (models 8 and 10). With this formulation, an increasing dilution factor modeled a decline in the magnitude of SGE with increasing group size, such that the magnitude of estimates for social effects was relevant to an average group size of 6.

We also proposed a customized (hereafter termed custom dilution) formulation to describe the change in estimates of SGE with group size for these data. This formulation was based on the proportional change to resources per sow with changes to group size. Values for Z4 were defined explicitly for each group size as the ratio of the count in the group relative to maximum pen capacity: Z4 = n/nmax. With this formulation, the social effects were thus assumed to decrease in magnitude proportional only due to reduction in the “fullness” of the pen.

All parameters were estimated using WOMBAT software (Meyer, 2006) and model comparisons were made using Akaike’s information criteria (AIC; Akaike, 1974). Significant changes resulting from dilution factors were evaluated by comparing log-likelihoods. All traits were treated as continuous due to the approximate normality of their distributions and the large number of values these traits could take.


RESULTS AND DISCUSSION

Data Augmentation

For experienced operators using appropriate equipment and procedures, ultrasonography is a reliable form of pregnancy detection in pigs (Kauffold et al., 1997). Therefore, although errors in pregnancy detection are possible, failed-to-farrow sows were assumed to be correctly assigned as pregnant before their grouping. In these data, 99.2% of sows that tested positive for pregnancy at around 28 to 35 d gestation were recorded to farrow, and 98.5% of sows produced at least 1 live piglet at farrowing. However, of the 1,827 groups of sows, only 1,747 groups (95.6%) contained all sows which farrowed, whereas the remaining 4.4% of groups included at least 1 sow that failed to farrow. Therefore, groups affected by a failure to farrow (4.4%) are overrepresented relative to the overall individual failure rate of 0.8%.

For sows that did not farrow (0.8% of records) data augmentation was required. Trait distributional properties were only slightly altered when comparing censored to augmented data (Table 2). However, it is necessary to have phenotypes for all animals in a group to estimate social effects. Other studies estimating indirect genetic effects do not appear to address the problem of data censoring within groups (e.g., due to mortality), although they may edit out otherwise incomplete groups (Bouwman et al., 2010). However, the most potentially informative animals (dead or removed from groups) will typically have no phenotype recorded, and therefore they do not contribute records toward the estimates of SGE of their group mates. Conversely, it is possible to estimate a social effect for an animal without its own phenotype. We suggest that the animals without “normal” farrowing phenotypes in these data were informative (i.e., their pregnancy failed in a group setting), so biologically meaningful phenotypes were substituted for missing values. The small deviation from normality in the distribution of trait values resulting from the data augmentation was not expected to influence the interpretation of results from this study. Data augmentation has also been performed in other studies where censoring has occurred (Johnston and Bunter, 1996).


View Full Table | Close Full ViewTable 2.

Raw data characteristics for augmented and censored data

 
Trait1 n Mean (SD) Minimum Maximum CV (×100%)
TB 10,748 11.41 (3.58) 0 26 31.4
TB censored 10,663 11.50 (3.45) 1 26 30.0
NBA 10,748 10.18 (3.44) 0 21 33.8
NBA censored 10,663 10.26 (3.33) 0 21 32.5
GL 10,748 116.9 (2.21) 73 130 1.9
GL censored 10,663 117.0 (1.79) 109 130 1.5
1TB, total born; NBA, number born alive; GL, gestation length.

The possible effects of a sow’s removal on the performance of their pen mates cannot be accurately accommodated in any model of analyses, and the extent of this in the data was not known. Sow removal without replacement has implications for the levels of sow disruption and space allowances during gestation. Sows that remain in the reduced static groups have improved space allowances, which might help offset earlier agonistic behavior from pen mates (if an aggressor has been removed). Alternatively, sows remaining in smaller groups (due to the removal of a victim) could be subjected to increased intensity of interactions with any remaining dominant sows, or hierarchy within the pen may need to be reestablished. The former would have no impact on TB, since fetal loss before a particular time point cannot be reversed. However, in the second case there may still be ongoing changes to sow weight, body condition, or fetal loss. It is not possible to know which situation, if either, occurred when sows were removed from pens. The possibilities for data censoring would be exacerbated in large dynamic gestation groups, and it is unclear if SGE could be estimated in this scenario.

Parameter Estimates

Estimates of parameters for TB, NBA, and GL under models ignoring group effects were similar to those reported in other studies for these traits recorded under stall housing (Tholen et al., 1996; Rothschild and Bidanel, 1998; Rydhmer et al., 2008). For TB, NBA, and GL, estimates of h2 were 0.11 ± 0.02, 0.12 ± 0.02, and 0.31 ± 0.02, with corresponding ratios for permanent environmental effects (pe2) of 0.11 ± 0.02, 0.10 ± 0.02, and 0.20 ± 0.02 (Table 3, model 1). Despite the relatively low proportion of sows with repeated records, the magnitude of estimates for permanent environmental effects were as expected for these traits. Moreover, estimates of variances and/or variance ratios for additive and SGE, or nongenetic group effects, were similar to those obtained from analyses that included only a single record per sow (i.e., analyzed excluding permanent environmental effects) using a restricted subset of this data (Bunter et al., 2014). Generally, estimates of h2 and pe2 were not significantly altered when models for analyses were expanded to include pen (model 1 vs. model 2) or group (model 1 vs. model 3) effects or SGE (model 1 vs. models 4 and 5).


View Full Table | Close Full ViewTable 3.

Model comparisons for parameter estimates assuming homogenous residual variance and zero covariance between additive and social genetic effects1

 
Model σ2a σ2pe σ2g σ2s σ2e σ2p h2 pe2 g2 s2 T2 LogL ΔAIC
Total born (pigs/litter)
1 1.281 1.225 ne ne 8.76 11.3 0.11 ± 0.02 0.11 ± 0.02 ne ne ne −18,279.70 15.5
2 1.280 1.222 0.0153 ne 8.75 11.3 0.11 ± 0.02 0.11 ± 0.02 0.001 ± 0.002 ne ne −18,279.46 17.1
3 1.266 1.203 0.156 ne 8.64 11.3 0.11 ± 0.02 0.11 ± 0.02 0.010 ± 0.006 ne ne −18,276.69 11.5
4 1.220 1.212 0.0416 0.0210 8.64 11.2 0.11 ± 0.02 0.11 ± 0.02 0.004 ± 0.007 0.002 ± 0.001 0.16 ± 0.02 −18,270.78 1.70
5 1.218 1.216 ne 0.0240 8.67 11.2 0.11 ± 0.02 0.11 ± 0.02 ne 0.002 ± 0.001 0.16 ± 0.02 −18,270.93 0
6 1.220 1.212 0.0421 0.0210 8.56 11.1 0.11 ± 0.02 0.11 ± 0.02 0.004 ± 0.007 0.002 ± 0.001 0.16 ± 0.03 −18,270.74 3.62
8.66 11.2 0.11 ± 0.02 0.11 ± 0.02 0.004 ± 0.007 0.002 ± 0.001 0.16 ± 0.03
Number born alive (pigs/litter)
1 1.185 1.003 ne ne 7.92 10.1 0.12 ± 0.02 0.10 ± 0.02 ne ne ne −17,698.19 10.2
2 1.184 1.003 0.0178 ne 7.90 10.1 0.12 ± 0.02 0.10 ± 0.02 0.002 ± 0.002 ne ne −17,697.81 11.5
3 1.168 0.985 0.1651 ne 7.78 10.1 0.12 ± 0.02 0.10 ± 0.02 0.020 ± 0.006 ne ne −17,694.10 4.04
4 1.132 0.994 0.0900 0.0140 7.79 10.1 0.11 ± 0.02 0.10 ± 0.02 0.009 ± 0.007 0.001 ± 0.001 0.15 ± 0.02 −17,691.24 0.32
5 1.127 1.002 ne 0.0201 7.84 10.1 0.11 ± 0.02 0.10 ± 0.02 ne 0.002 ± 0.001 0.16 ± 0.02 −17,692.08 0
6 1.132 0.993 0.0885 0.0144 8.03 10.3 0.11 ± 0.02 0.10 ± 0.02 0.009 ± 0.007 0.001 ± 0.001 0.14 ± 0.02 −17,690.84 1.52
7.73 10.0 0.11 ± 0.02 0.10 ± 0.02 0.009 ± 0.007 0.001 ± 0.001 0.15 ± 0.03
Gestation length (days)
1 1.270 0.812 ne ne 2.07 4.15 0.31 ± 0.02 0.20 ± 0.03 ne ne ne −12,384.43 20.6
2 1.270 0.812 0.0002 ne 2.07 4.15 0.31 ± 0.02 0.20 ± 0.03 0.000 ± 0.002 ne ne −12,384.43 22.6
3 1.257 0.818 0.0535 ne 2.02 4.15 0.30 ± 0.02 0.20 ± 0.03 0.013 ± 0.005 ne ne −12,380.99 15.7
4 1.229 0.831 0.0198 0.0069 2.02 4.13 0.30 ± 0.02 0.20 ± 0.03 0.005 ± 0.006 0.002 ± 0.001 0.34 ± 0.03 −12,376.35 8.46
5 1.228 0.833 ne 0.0084 2.02 4.13 0.30 ± 0.02 0.20 ± 0.03 ne 0.002 ± 0.001 0.35 ± 0.03 −12,376.66 7.08
6 1.216 0.819 0.0259 0.0060 1.73 3.82 0.32 ± 0.03 0.21 ± 0.03 0.007 ± 0.007 0.002 ± 0.001 0.36 ± 0.03 −12,371.12 0
2.10 4.19 0.29 ± 0.02 0.20 ± 0.02 0.006 ± 0.006 0.001 ± 0.001 0.33 ± 0.03
1σ2a, additive genetic variance; σ2pe, permanent environmental variance; σ2g, group variance (group membership is defined by pen in model 2 and pen × gestation group in model 3); σ2s, variance due to social genetic effects; σ2e, residual variance; σ2p, phenotypic variance; h2 = σ2a/σ2p; pe2 = σ2pe/σ2p; g2 = σ2g/σ2p; s2 = σ2s/σ2p; T2 = [σ2a +(6 − 1) 2σ2s]/σ2p; LogL, log-likehood; ΔAIC, change in Akaike’s information criteria from the best (minimum) model AIC; ne, not estimated; parameters are presented using both levels of residuals in calculations for model 6.

Estimates for nongenetic pen effects were not significant for any reproductive trait (Table 3, model 1 vs. model 2). This provides evidence that characteristics of individual gestation pens per se have no significant impact on the reproductive performance of sows. For a pregnancy well established before grouping of sows for gestation, as was the case here, it was not expected that the location of the sow or group within a gestation barn (i.e., defined by pen) would have a direct influence on reproductive outcomes. Moreover, the resources supplied per pen were essentially identical at this facility, in terms of the space allowed per sow and their construction and design, with respect to feeders, ventilation, and light.

In contrast to results for pen, estimates of nongenetic group effects (i.e., defined by grouping within pen) were associated with significant improvements to the models for analysis for all traits (model 3 vs. model 1) even though variation due to nongenetic group was very low. Proportional variance due to group effects (g2) for reproductive traits ranged from 0.01 to 0.02 (Table 3), in contrast to typically higher estimates (0.17 to 0.27) obtained for growth traits (Bergsma et al., 2008; Jones et al., 2011). However, the ratio of group variance to additive genetic variance ranged from 9 to 17% for reproductive traits. In the absence of significant nongenetic pen effects for reproductive traits and with parity and breed already accounted for in models for analyses, the random nongenetic group term more convincingly represents variation resulting from within group factors, such as interactions between sows. Thus, obtaining further significant improvements in model fit by adding terms for SGE (models 4 and 5, Table 3) was expected. However, based on the change in AIC (ΔAIC; Table 3), fitting a model including SGE without an additional term for nongenetic pen or group effects (model 5) was slightly better than model 4 for all traits. This implies that environmental covariance between individuals within the group was low. However, residual variance and the variance due to social effects were also lower under model 4 compared to model 5 for TB and NBA. Estimates of SGE can be biased upward by environmental covariance between group mates when the term for the nongenetic group is not fitted (Muir, 2005; Bijma, 2010b).

The estimated proportional variance of SGE (s2 ) was very small, ranging from 0.001 to 0.002 (Table 3). This low variance is consistent with estimates obtained elsewhere for other traits, such as growth, where variation between groups is typically also much larger (Arango et al., 2005; Chen et al., 2007; Bergsma et al., 2008). However, the contribution of SGE to total genetic variation for reproductive traits was significant. Ratios of total genetic variance to phenotypic variance (T2) were 0.16 ± 0.02 for TB and ranged between 0.14 ± 0.02 to 0.16 ± 0.02 for NBA or 0.33 ± 0.02 to 0.36 ± 0.02 for GL (Table 3). Compared to h2 variation, the total genetic variation (T2) available for selection purposes was increased by approximately 33, 25, and 12% for TB, NBA, and GL using estimates from models accounting for SGE. When the covariance between additive and social effects was also estimated (Table 4, dilution 0), values for T2 were further increased to 0.20 ± 0.02 for TB and NBA and 0.42 ± 0.04 for GL.


View Full Table | Close Full ViewTable 4.

Parameter estimates from models fitting different dilution factors1

 
Dilution Model σ2a σ2pe σ2s σas σ2e σ2p h2 pe2 s2 T2 ras ΔAIC
Total born (pigs/litter)
0 7 1.229 1.220 0.017 0.056 8.69 11.3 0.11 ± 0.02 0.11 ± 0.02 0.002 ± 0.001 0.20 ± 0.03 0.38 ± 0.22 29.2
0.2 8 1.226 1.221 0.020 0.055 8.68 11.3 0.11 ± 0.02 0.11 ± 0.02 0.002 ± 0.001 0.20 ± 0.03 0.35 ± 0.21 29.5
0.4 8 1.225 1.222 0.021 0.053 8.68 11.3 0.11 ± 0.02 0.11 ± 0.02 0.002 ± 0.001 0.20 ± 0.03 0.32 ± 0.19 30.1
0.6 8 1.226 1.222 0.023 0.049 8.68 11.3 0.11 ± 0.02 0.11 ± 0.02 0.002 ± 0.001 0.20 ± 0.03 0.29 ± 0.19 31.4
0.8 8 1.230 1.221 0.022 0.045 8.68 11.3 0.11 ± 0.02 0.11 ± 0.02 0.002 ± 0.001 0.20 ± 0.03 0.27 ± 0.19 33.5
1 8 1.237 1.219 0.019 0.040 8.68 11.3 0.11 ± 0.02 0.11 ± 0.02 0.002 ± 0.001 0.19 ± 0.03 0.26 ± 0.19 36.7
Custom 8 1.201 1.306 0.059 0.002 8.43 11.2 0.11 ± 0.02 0.12 ± 0.02 0.005 ± 0.001 0.24 ± 0.03 0.00 ± 0.12 0
Number born alive (pigs/litter)
0 7 1.134 1.007 0.013 0.053 7.87 10.1 0.11 ± 0.02 0.10 ± 0.02 0.001 ± 0.001 0.20 ± 0.03 0.44 ± 0.25 28.4
0.2 8 1.132 1.007 0.015 0.053 7.87 10.1 0.11 ± 0.02 0.10 ± 0.02 0.001 ± 0.001 0.20 ± 0.03 0.41 ± 0.23 28.6
0.4 8 1.130 1.007 0.016 0.052 7.86 10.1 0.11 ± 0.02 0.10 ± 0.02 0.002 ± 0.001 0.20 ± 0.03 0.38 ± 0.22 29.1
0.6 8 1.130 1.006 0.017 0.050 7.86 10.1 0.11 ± 0.02 0.10 ± 0.02 0.002 ± 0.001 0.20 ± 0.03 0.36 ± 0.21 30.1
0.8 8 1.133 1.004 0.017 0.047 7.86 10.1 0.11 ± 0.02 0.10 ± 0.02 0.002 ± 0.001 0.20 ± 0.03 0.34 ± 0.21 31.8
1 8 1.139 1.002 0.015 0.043 7.87 10.1 0.11 ± 0.02 0.10 ± 0.02 0.001 ± 0.001 0.19 ± 0.03 0.33 ± 0.22 34.2
Custom 8 1.101 1.107 0.056 0.0002 7.59 10.1 0.11 ± 0.02 0.11 ± 0.02 0.006 ± 0.001 0.25 ± 0.03 0.00 ± 0.12 0
Gestation length (d)2
0 9 1.231 0.835 0.006 0.024 1.76 3.85 0.32 ± 0.03 0.21 ± 0.03 0.001 ± 0.001 0.42 ± 0.04 0.31 ± 0.19 0.32
0.2 10 1.217 0.823 0.007 0.026 1.75 3.85 0.32 ± 0.03 0.21 ± 0.03 0.002 ± 0.001 0.42 ± 0.04 0.29 ± 0.17 0
0.4 10 1.215 0.824 0.008 0.025 1.74 3.85 0.32 ± 0.03 0.21 ± 0.03 0.002 ± 0.001 0.43 ± 0.04 0.26 ± 0.17 0.06
0.6 10 1.215 0.823 0.008 0.025 1.74 3.84 0.32 ± 0.03 0.21 ± 0.03 0.002 ± 0.001 0.43 ± 0.04 0.25 ± 0.16 0.8
0.8 10 1.218 0.821 0.008 0.025 1.73 3.83 0.32 ± 0.03 0.21 ± 0.03 0.002 ± 0.001 0.43 ± 0.04 0.27 ± 0.16 2.64
1 10 1.226 0.816 0.006 0.026 1.73 3.83 0.32 ± 0.03 0.21 ± 0.03 0.001 ± 0.001 0.43 ± 0.04 0.32 ± 0.18 5.64
Custom 10 1.231 0.823 0.006 0.024 2.03 4.16 0.30 ± 0.02 0.20 ± 0.02 0.001 ± 0.001 0.39 ± 0.04 0.28 ± 0.18 10.7
1σ2a, additive genetic variance; σ2pe, permanent environmental variance; σ2s, variance due to social genetic effects; sas, covariance between additive and social genetic effects; σ2e, residual variance; σ2p, phenotypic variance; h2 = σ2a/σ2p; pe2 = σ2pe/σ2p; s2 = σ2s/σ2p; T2 = [σ2a + (6 − 1) 2σ2s]/σ2p; ras, correlation between additive and social genetic effects; ΔAIC, change in Akaike’s information criteria from the best (minimum) model AIC.
2Variances and parameter estimates presented for group size class ≤4 only; entries of 0.00 indicate that the value was ≤0.01.

Heterogeneous Variances

Fitting heterogeneous variances without dilution did not improve model fit for TB or NBA, but it did significantly improve the model for GL (Table 3, model 6). This was probably because a single extreme value for GL in a group of 4 sows has a significantly bigger impact on variation within the group than a single extreme value in a larger group. Since extreme values for GL were only present due to pregnancy loss, it seems unlikely that the estimates obtained for GL reflect a social genetic effect on gestation length for otherwise normal pregnancies. Bijma (2010b) suggested that to obtain unbiased estimates for SGE relevant to the average group size, it may be required to fit separate group and residual variance for each group size. However, this is hard to achieve in practice, and there was limited evidence for heterogeneity in variance for litter size resulting from group size. The addition of nongenetic group effects or heterogeneous variances fitted in combination with dilution factors or both (i.e., extensions of models 8 and 10) did not further improve model fit for any trait and did not significantly alter any parameter estimates for this data (not presented).

The Impact of Dilution Factors on Parameter Estimates

Dilution factors applied to the linear functions of group size (i.e., [( − 1)/(n − 1)]d) resulted in no significant improvement in the log-likelihood for any of the traits analyzed. Conversely, the fit of the model to the data gradually deteriorated as the dilution factor increased for TB and NBA (Table 4), implying that varying the magnitude of SGE with this function was not a suitable strategy. For GL, the best model fit occurred with a dilution of 0.2 but significantly deteriorated for dilution factors of 0.8 and 1. There was no convincing evidence for a decline in the magnitude of estimates for social effects on litter size as sow numbers within a group increased above 6 or, concurrently, an increase in the magnitude of SGE when there were fewer than 6 sows per group. However, 79% of records in this data come from groups of size 6, 7, and 8, so the dilution factors also resulted in relatively small changes to the coefficients of Z4 at these pen sizes. Dilution of SGE with increasing group size, originally proposed by Bijma (2010b), probably better describes outcomes when resources per group, and not per individual, are fixed because of the implications for competitive interactions.

In contrast, allowing estimates of social effects to vary proportional to n/nmax gave a substantial improvement in the model fit for TB and NBA, but not GL, as indicated by an improved AIC (custom dilution, Table 4). Coefficients of Z4 were lower for groups where n < nmax, which implies that the magnitude of social effects was reduced in all pen sizes when fewer sows were held in groups relative to nmax. In this data, the distribution of resources per sow in terms of space and feed delivery were constant regardless of nmax. That is, whether nmax was 4, 8, or 10, the space per sow and the feed delivered per sow were identical where n = nmax, but the number of sows to interact with and pen dimensions differed. Where n < nmax, then both the number of sows to interact with was reduced and the space allocation per sow was increased proportional to the reduction in n relative to nmax, while pen dimensions and feed delivery/sow were not altered within a specific nmax. Therefore, the ratio of n/nmax essentially described both the proportional reduction in sows to interact with and an increase in area per sow. Simple calculations show that the ratio of change in the number of sows available for interactions (n − 1/nmax − 1) produces similar coefficients to the ratio describing changes in area (n/nmax), particularly as n becomes larger.

The different results arising from alternative dilution factors highlight the importance of having a sensible model to describe the likely magnitude of social effects in real data and the implications of group size for these effects. An increasing dilution of SGE with increasing group size did not improve model fit for reproductive traits. In contrast, allowing a reduction in estimates of SGE proportional to decreasing group size relative to the space allowance was more appropriate for TB and NBA. This model (custom dilution, Table 4) also resulted in estimates of correlations between additive and SGE close to zero.

The Covariance Between Additive and SGE

When there was no dilution of social effect with increasing group size, the correlations between additive and SGE were positive in this study (Table 4). A positive correlation implies that a sow’s additive merit for her own reproduction was positively associated with her contribution to the reproductive performance of her pen mates. Trubenova and Hagar (2012) suggested that it is difficult to think of a scenario where genes that are good for an animal’s own performance are also good for its group mates in a competitive situation. Restrictive feeding during gestation typically occurs in both nucleus and commercial populations, introducing a competitive element for feed. This argument therefore suggests that a positive correlation between additive and SGE is generally unlikely. Moreover, a positive correlation was not observed in a previous study using univariate subsets of this data, where the correlation between estimates of additive and social effects was slightly negative (Bunter et al., 2014). Based on the change in AIC from the best model per trait (Table 3) and dilution 0 (Table 4), the fit of the model for TB and NBA but not GL was significantly (P < 0.05) improved by allowing a covariance between direct and SGE, although SE of estimates were large. However, models with the best fit for TB and NBA (custom dilution, Table 4) ultimately resulted in an estimate of the correlation between direct and SGE no different to zero. Based on combined results from this and previous (Bunter et al., 2014) analyses, it seems most likely that the genetic covariance between direct and SGE is low for reproductive traits. Moreover, the estimated covariance between additive and SGE was sensitive to the modeling for SGE with respect to group size. For all traits, there were no significant changes to estimates with the addition of nongenetic group effects or heterogeneous variances fitted in combination with dilution factors or both (not presented).

A low to 0 correlation between social genetic and additive genetic effects can also be supported by the argument that these estimates do not actually represent the same trait expression. From a biological perspective, it is not the individual expression of TB itself, which is a function of ovulation rate and embryo survival, that directly affected the TB of pen mates. The estimated social effect for an individual is either a consequence of unobserved traits expressed during gestation (e.g., aggression, monopolization of resources) or an indirect effect on pen mates for a trait correlated with TB, which then influences phenotypes of pen mates for TB. In this situation, it seems unlikely that genes that affect the expression of aggression, for example, are the same genes that affect ovulation rate and embryo survival, but there are no data to confirm this. This is in contrast to a trait like growth rate, where the actual growth rate of an individual can directly affect the growth rate of other individuals by altering ranking for competitive advantage at the feeder, thereby resulting in a negative genetic correlation between direct and SGE, as described by Trubenova and Hagar (2012).

Practical Implications for Genetic Evaluation

Previous work (Bunter et al., 2014) demonstrated that SGE affected expression of reproductive traits for group-housed sows, were repeatable across different sow groupings, and improved the prediction of mean group performance. This implies that reproductive performance and welfare of sows housed in groups might also be improved through considering SGE in genetic evaluation systems. In this case, the mechanism most likely is an unfavorable relationship between aggression of an individual and the reproduction of its pen mates, but data to confirm this are generally not available. In the presence of estimable SGE, response to selection might be improved through the inclusion of both additive and SGE into a selection index (Muir, 2005).

More generally, associations between SGE and behaviors relating to welfare, as well as the implications of an animal’s behavior for its own performance, need to be better quantified (Cassady, 2007) before it will be widely accepted that competitive effects models have a useful place for specific traits in large animal breeding applications. In contrast to the clearly demonstrated results from selection experiments in poultry (Muir, 1996a,b) or quail (Muir 2005), where the impact of mortality resulting from aggression on performance is significant, results for selection experiments involving growing pigs have been more ambiguous. Progeny groups of pigs divergently selected for SGE estimated from growth data showed no significant differences in growth rate, measures of aggression, or the time spent fighting but demonstrated significantly less biting behavior in the group of animals with favorable social effects (Camerlink et al., 2010, 2013, 2014, 2015). Moreover, estimates of correlations between additive or SGE for growth and observed fighting or bullying behavior, or fight lesions, were shown to depend on the correlation between additive and SGE. A wide range in the assumed underlying genetic correlation between additive and SGE (−0.58 to +0.58) was feasible for phenotypes observed (Canario et al., 2012). This could suggest that in addition to lack of a response in performance traits, welfare benefits will also not always be observed under selection solely for SGE.

It is also important to consider the implications of missing phenotypes within groups when estimating initial parameters for SGE and for estimating breeding values for individuals. Bouwman et al. (2010) suggested that a missing own phenotype for individual, i, in group j was not a limitation to estimating the social genetic effect of individual i and was therefore unproblematic. We argue, however, that missing phenotype(s) can have implications for estimating the social effects for group mates of i in group j, assuming that at least some missing phenotypes are not at random with respect to the unrecorded value of the indicator trait. That is, nonrandomly censored data needs to be included in analyses, or the impact of social effects and the magnitude of social breeding values may be underestimated. The same is true for other traits, like growth, where it can be speculated that an unrecorded animal removed from its group typically has a missing phenotype that was potentially prefaced by poor growth as a consequence of aggressive behavior toward it. The ability to identify phenotype censoring within groups and the appropriate data augmentation to create meaningful proxy phenotypes requires well-structured and complete recording systems where group membership and reasons for censoring are known or can be accurately inferred.

The introduction of group housing for gestating sows ultimately has implications for their reproductive performance primarily because it allows nonbeneficial interactions to occur between sows. The results from this and a previous (Bunter et al., 2014) study suggest that there is a heritable component to social interactions, which are significantly affecting reproductive performance of pen mates for group-housed sows. This result occurred even where housing design and grouping strategies (Marchant-Forde and Marchant-Forde, 2005; Li et al., 2012; Gonyou and Rioja-Lang, 2014) were implemented to reduce outcomes from negative social interactions and where the impact of early embryonic loss was not represented in phenotypes as a consequence of group housing because sows were individually stalled before about 35 d of gestation. It may be speculated that in herds where grouping strategies do not minimize the likelihood of detrimental interactions, SGE on reproductive traits could be larger if the incidence of embryonic or pregnancy loss also increased. Moreover, grouping sows at weaning or mixing around the more sensitive time of embryo implantation, as currently occurs in the United Kingdom and Australia, might also increase estimates of SGE for reproductive traits due to potential contributions of social effects to early embryonic losses. Therefore, this study might underestimate the magnitude of SGE in such systems.

The best models for litter size traits in this data allowed estimates of social effects to reduce in magnitude as more space became available per sow but did not dilute the magnitude of social effects for full pens of 4, 8, or 10 sows/pen, which provide constant space per sow in static groups. This analysis therefore supports results from other studies on sow grouping, which show improved reproductive performance with an increased space allowance per sow (Hemsworth et al., 2013). This outcome also implies that the number of pen mates present did not alter the magnitude of SGE when the pen was full. However, the increase in pen dimensions with increasing group size was modest (changing from 2.44 to 4.88 m on the long side for groups of 4 and 8) such that animals are unlikely to be able to escape from the personal space of an antagonist regardless of group size in this data. Results also suggest that in housing systems where resources supplied per sow are effectively equal and groups are already structured to reduce the incidence of aggression between pen mates, the absolute group size was relatively less important at least with groups of ≤10 sows. However, group size might be more influential on performance when resources per sow are not as well controlled or with dramatically different group sizes (e.g., 10 vs. 40 sows per pen), which concurrently provide changes to pen dimensions enabling effective escape from protagonists.

A complicating factor for estimating SGE for any trait is the issue of variable group size and the utility of such estimates for groups of different size (Bijma, 2010b). Interactions between sows are not necessarily competitive, so the number of sows to interact with is mostly pertinent when interactions are not neutral. Moreover, as noted above, group size is confounded with pen dimensions, which also has implications for a sow’s ability to escape antagonists. These factors may create an imperfect relationship between group size and the magnitude of SGE, which are estimated from real data. Gestation groups in this data were also relatively small by commercial standards, but SGE will be much harder to estimate in large groups because of the rapid increase in the number of possible interactions between animals (Muir 2005). Moreover, gestation groups can be dynamic, and it is unclear whether SGE could be estimated in this scenario at all. This certainly provides motivation for investigating alternative sources of group data to estimate SGE, such as earlier growth data in pigs. Moreover, since SGE are estimated as indirect effects from performance data and are therefore expressed in the scale of the performance trait used, a universal scale to transform such estimates between traits would be required, if that is possible. Use of different data sources implies that the estimates of social effects for different traits could be due to a common, unobserved, phenotype, which affects more than 1 trait recorded in social cohorts (e.g., aggressive behavior affecting both growth of finishers and, later, reproductive performance of sows). For practical application, more work is required to obtain meaningful ranking of animals across a range of group sizes and to establish how different performance traits might provide data on SGE. Additional work is also required to incorporate social effect models into genetic evaluation systems where genomic data are also used.

It can also be questioned whether the results from highly parameterized models, such as those necessary to estimate SGE, are more likely to be spurious. The data used here to estimate SGE for reproductive data were not of optimal design, but according to Bijma (2010a), the number of records and groups present should be sufficient to estimate the parameters with reasonable accuracy. Moreover, use of a conservative AIC to compare alternative models and allowing for the implications of parameters close to the boundary of the parameter space on significance testing (Visscher, 2006) somewhat supports the concept that the estimates for SGE were not spurious, even if the underlying mechanism was not known or illustrated by concurrent data on sow behavior. Nevertheless, implementation of more highly parameterized models in genetic evaluation systems can be complicated and also requires adequate data recording to be effective.

Unfortunately, no model used in this study allowed for sows to interact differentially with other individual sows within their group, which is also more likely than the underlying assumption that each sow has an identical effect on the phenotypes of all pen mates. Observation of individual sows shows that they do not have identical interactions with all pen mates because a hierarchy is present within groups (Brouns and Edwards, 1994). Therefore, existing social effects models are averaging the impact of an individual on pen mates across all animals within the group, when a model formulated to accommodate hierarchical effects might be more effective.

 

References

Footnotes


Comments
Be the first to comment.



Please log in to post a comment.
*Society members, certified professionals, and authors are permitted to comment.