Evaluation of measures of correctness of genotype imputation in the context of genomic prediction: a review of livestock applications

In livestock, many studies have reported the results of imputation to 50k single nucleotide polymorphism (SNP) genotypes for animals that are genotyped with low-density SNP panels. The objective of this paper is to review different measures of correctness of imputation, and to evaluate their utility depending on the purpose of the imputed genotypes. Across studies, imputation accuracy, computed as the correlation between true and imputed genotypes, and imputation error rates, that counts the number of incorrectly imputed alleles, are commonly used measures of imputation correctness. Based on the nature of both measures and results reported in the literature, imputation accuracy appears to be a more useful measure of the correctness of imputation than imputation error rates, because imputation accuracy does not depend on minor allele frequency (MAF), whereas imputation error rate depends on MAF. Therefore imputation accuracy can be better compared across loci with different MAF. Imputation accuracy depends on the ability of identifying the correct haplotype of a SNP, but many other factors have been identified as well, including the number of genotyped immediate ancestors, the number of animals with genotypes at the high-density panel, the SNP density on the low- and high-density panel, the MAF of the imputed SNP and whether imputed SNP are located at the end of a chromosome or not. Some of these factors directly contribute to the linkage disequilibrium between imputed SNP and SNP on the low-density panel. When imputation accuracy is assessed as a predictor for the accuracy of subsequent genomic prediction, we recommend that: (1) individual-specific imputation accuracies should be used that are computed after centring and scaling both true and imputed genotypes; and (2) imputation of gene dosage is preferred over imputation of the most likely genotype, as this increases accuracy and reduces bias of the imputed genotypes and the subsequent genomic predictions.     

Efficiency of genomic prediction for boar taint reduction in Danish Landrace pigs

Genetic selection against boar taint, which is caused by high skatole and androstenone concentrations in fat, is a more acceptable alternative than is the current practice of castration. Genomic predictors offer an opportunity to overcome the limitations of such selection caused by the phenotype being expressed only in males at slaughter, and this study evaluated different approaches to obtain such predictors. Samples from 1000 pigs were included in a design which was dominated by 421 sib pairs, each pair having one animal with high and one with low skatole concentration (≥0.3 μg/g). All samples were measured for both skatole and androstenone and genotyped using the Illumina SNP60 porcine BeadChip for 62 153 single nucleotide polymorphisms. The accuracy of predicting phenotypes was assessed by cross‐validation using six different genomic evaluation methods: genomic best linear unbiased prediction (GBLUP) and five Bayesian regression methods. In addition, this was compared to the accuracy of predictions using only QTL that showed genome‐wide significance. The range of accuracies obtained by different prediction methods was narrow for androstenone, between 0.29 (Bayes Lasso) and 0.31 (Bayes B), and wider for skatole, between 0.21 (GBLUP) and 0.26 (Bayes SSVS). Relative accuracies, corrected for h², were 0.54–0.56 and 0.75–0.94 for androstenone and skatole respectively. The whole‐genome evaluation methods gave greater accuracy than using only the QTL detected in the data. The results demonstrate that GBLUP for androstenone is the simplest genomic technology to implement and was also close to the most accurate method. More specialised models may be preferable for skatole.     

Toward genomic prediction from whole-genome sequence data: impact of sequencing design on genotype imputation and accuracy of predictions

Genomic prediction from whole-genome sequence data is attractive, as the accuracy of genomic prediction is no longer bounded by extent of linkage disequilibrium between DNA markers and causal mutations affecting the trait, given the causal mutations are in the data set. A cost-effective strategy could be to sequence a small proportion of the population, and impute sequence data to the rest of the reference population. Here, we describe strategies for selecting individuals for sequencing, based on either pedigree relationships or haplotype diversity. Performance of these strategies (number of variants detected and accuracy of imputation) were evaluated in sequence data simulated through a real Belgian Blue cattle pedigree. A strategy (AHAP), which selected a subset of individuals for sequencing that maximized the number of unique haplotypes (from single-nucleotide polymorphism panel data) sequenced gave good performance across a range of variant minor allele frequencies. We then investigated the optimum number of individuals to sequence by fold coverage given a maximum total sequencing effort. At 600 total fold coverage (x 600), the optimum strategy was to sequence 75 individuals at eightfold coverage. Finally, we investigated the accuracy of genomic predictions that could be achieved. The advantage of using imputed sequence data compared with dense SNP array genotypes was highly dependent on the allele frequency spectrum of the causative mutations affecting the trait. When this followed a neutral distribution, the advantage of the imputed sequence data was small; however, when the causal mutations all had low minor allele frequencies, using the sequence data improved the accuracy of genomic prediction by up to 30%.     

Use of field data in pig genomic selection schemes: a simulation study

The aim of this study was to test how genetic gain for a trait not measured on the nucleus animals could be obtained within a genomic selection pig breeding scheme. Stochastic simulation of a pig breeding program including a breeding nucleus, a multiplier to produce and disseminate semen and a production tier where phenotypes were recorded was performed to test (1) the effect of obtaining phenotypic records from offspring of nucleus animals, (2) the effect of genotyping production animals with records for the purpose of including them in a genomic selection reference population or (3) to combine the two approaches. None of the tested strategies affected genetic gain if the trait under investigation had a low economic value of only 10% of the total breeding goal. When the relative economic weight was increased to 30%, a combination of the methods was most effective. Obtaining records from offspring of already genotyped nucleus animals had more impact on genetic gain than to genotype more distant relatives with phenotypes to update the reference population. When records cannot be obtained from offspring of nucleus animals, genotyping of production animals could be considered for traits with high economic importance.     

Accuracy and bias of genomic prediction with different de-regression methods

Genomic selection has become increasingly important in the breeding of animals and plants. The response variable is an important factor, influencing the accuracy of genomic selection. The de-regressed proof (DRP) based on traditional estimated breeding value (EBV) is commonly used as response variable. In the current study, simulated data from 16th QTL-MAS Workshop and real data from Chinese Holstein cattle were used to compare accuracy and bias of genomic prediction with two methods of calculating DRP. Our results with simulated data showed that the correlation between genomic EBV and true breeding value achieved using the Jairath method (DRP_J) was superior to that achieved using the Garrick method (DRP_G) for simulated trait 1 but the reverse was true for simulated trait 3, and these two methods performed comparably for simulated trait 2. For all three simulated traits, DRP_J yielded larger bias of genomic prediction. However, DRP_J outperformed DRP_G in both accuracy and unbiasedness for four milk production traits in Chinese Holstein. In the estimation of genomic breeding value using genomic BLUP model, two methods for weighting diagonal elements of incidence matrix associated with residual error were also compared. With increasing the proportion of genetic variance unexplained by markers, the accuracy of genomic prediction was decreased and the bias was increased. Weighting by the reliability of DRP produced accuracy comparable to the evaluation where the proportion of genetic variance unexplained by markers was considered, but with smaller bias in general.     

The importance of phenotypic data analysis for genomic prediction - a case study comparing different spatial models in rye

Background

Genomic prediction is becoming a daily tool for plant breeders. It makes use of genotypic information to make predictions used for selection decisions. The accuracy of the predictions depends on the number of genotypes used in the calibration; hence, there is a need of combining data across years. A proper phenotypic analysis is a crucial prerequisite for accurate calibration of genomic prediction procedures. We compared stage-wise approaches to analyse a real dataset of a multi-environment trial (MET) in rye, which was connected between years only through one check, and used different spatial models to obtain better estimates, and thus, improved predictive abilities for genomic prediction. The aims of this study were to assess the advantage of using spatial models for the predictive abilities of genomic prediction, to identify suitable procedures to analyse a MET weakly connected across years using different stage-wise approaches, and to explore genomic prediction as a tool for selection of models for phenotypic data analysis.

Results

Using complex spatial models did not significantly improve the predictive ability of genomic prediction, but using row and column effects yielded the highest predictive abilities of all models. In the case of MET poorly connected between years, analysing each year separately and fitting year as a fixed effect in the genomic prediction stage yielded the most realistic predictive abilities. Predictive abilities can also be used to select models for phenotypic data analysis. The trend of the predictive abilities was not the same as the traditionally used Akaike information criterion, but favoured in the end the same models.

Conclusions

Making predictions using weakly linked datasets is of utmost interest for plant breeders. We provide an example with suggestions on how to handle such cases. Rather than relying on checks we show how to use year means across all entries for integrating data across years. It is further shown that fitting of row and column effects captures most of the heterogeneity in the field trials analysed.

Electronic supplementary material

The online version of this article (doi:10.1186/1471-2164-15-646) contains supplementary material, which is available to authorized users.     

Comparison of accuracy of intramuscular fat prediction in live pigs using five different ultrasound intensity levels

The objective of this study was to evaluate the possibility of prediction of intramuscular fat (IMF) in live pigs using ultrasound method. Moreover, the accuracy of prediction at five different ultrasound intensity levels was investigated. Cross-sectional images of longissimus dorsi muscle (LD) at right last rib area, from hybrid pigs, were taken. Each pig was scanned at the same frequency (3.5 MHz) and at the five different ultrasound intensity levels 70%, 75%, 80%, 85% and 90% of total amplifying of sonograph, using the device ALOKA SSD-500. The video image analysis was used to predict IMF content (ultrasound intramuscular fat (UIMF) 70 to UIMF90). The second day after slaughter, the dissection of right half carcass was done. A sample of LD at the last rib was taken for laboratory analysis of IMF content (LAIMF). Scatter plots with UIMF on the x-axis and LAIMF on the y-axis were constructed to account for individual variability within and between intensity levels. Correlations between LAIMF and UIMF were significantly different from zero (r = 0.40–0.52), except for correlation between LAIMF and UIMF90 (r = 0.14). Statistical model with LAIMF (the dependent variable), UIMF (the same model for each intensity level), live weight (the covariates) and sex (the fixed effect) was developed. Coefficients of determination (R²) were 0.33, 0.38, 0.34, 0.25 and 0.17 with UIMF at the intensity level 70%, 75%, 80%, 85% and 90%. Root mean square errors ranged from 0.516% to 0.639%. Standard errors of individual prediction ranged from 0.523% to 0.649%. Goodness-of-fit of the model was also justified by testing the residuals for normality. Although the results are not quite unequivocal in favour of the one intensity level, it seems that intensity levels 75% and 80% are the most suitable to predict IMF in live pigs. Further research is needed, mainly to increase accuracy of collecting, processing and evaluating the sonograms using video image analysis.     

Impact of genotyping strategy on the accuracy of genomic prediction in simulated populations of purebred swine

Single-step genomic BLUP (ssGBLUP) has been widely used in genomic evaluation due to relatively higher prediction accuracy and simplicity of use. The prediction accuracy from ssGBLUP depends on the amount of information available concerning both genotype and phenotype. This study investigated how information on genotype and phenotype that had been acquired from previous generations influences the prediction accuracy of ssGBLUP, and thus we sought an optimal balance about genotypic and phenotypic information to achieve a cost-effective and computationally efficient genomic evaluation. We generated two genetically correlated traits (h² = 0.35 for trait A, h² = 0.10 for trait B and genetic correlation 0.20) as well as two distinct populations mimicking purebred swine. Phenotypic and genotypic information in different numbers of previous generations and different genotyping rates for each litter were set to generate different datasets. Prediction accuracy was evaluated by correlating genomic estimated breeding values with true breeding values for genotyped animals in the last generation. The results revealed a negligible impact of previous generations that lacked genotyped animals on the prediction accuracy. Phenotypic and genotypic data, including the most recent three to four generations with a genotyping rate of 40% or 50% for each litter, could lead to asymptotic maximum prediction accuracy for genotyped animals in the last generation. Single-step genomic best linear unbiased prediction yielded an optimal balance about genotypic and phenotypic information to ensure a cost-effective and computationally efficient genomic evaluation of populations of polytocous animals such as purebred pigs.     

Repurposing population genetics data to discern genomic architecture: A case study of linkage cohort detection in mountain pine beetle (Dendroctonus ponderosae)

Genetic surveys of the population structure of species can be used as resources for exploring their genomic architecture. By adjusting filtering assumptions, genome‐wide single‐nucleotide polymorphism (SNP) datasets can be reused to give new insights into the genetic basis of divergence and speciation without targeted resampling of specimens. Filtering only for missing data and minor allele frequency, we used a combination of principal components analysis and linkage disequilibrium network analysis to distinguish three cohorts of variable SNPs in the mountain pine beetle in western Canada, including one that was sex‐linked and one that was geographically associated. These marker cohorts indicate genomically localized differentiation, and their detection demonstrates an accessible and intuitive method for discovering potential islands of genomic divergence without a priori knowledge of a species’ genomic architecture. Thus, this method has utility for directly addressing the genomic architecture of species and generating new hypotheses for functional research.     

Accuracy of genomic prediction within and across populations for nematode resistance and body weight traits in sheep

Genomic prediction utilizes single nucleotide polymorphism (SNP) chip data to predict animal genetic merit. It has the advantage of potentially capturing the effects of the majority of loci that contribute to genetic variation in a trait, even when the effects of the individual loci are very small. To implement genomic prediction, marker effects are estimated with a training set, including individuals with marker genotypes and trait phenotypes; subsequently, genomic estimated breeding values (GEBV) for any genotyped individual in the population can be calculated using the estimated marker effects. In this study, we aimed to: (i) evaluate the potential of genomic prediction to predict GEBV for nematode resistance traits and BW in sheep, within and across populations; (ii) evaluate the accuracy of these predictions through within-population cross-validation; and (iii) explore the impact of population structure on the accuracy of prediction. Four data sets comprising 752 lambs from a Scottish Blackface population, 2371 from a Sarda×Lacaune backcross population, 1000 from a Martinik Black-Belly×Romane backcross population and 64 from a British Texel population were used in this study. Traits available for the analysis were faecal egg count for Nematodirus and Strongyles and BW at different ages or as average effect, depending on the population. Moreover, immunoglobulin A was also available for the Scottish Blackface population. Results show that GEBV had moderate to good within-population predictive accuracy, whereas across-population predictions had accuracies close to zero. This can be explained by our finding that in most cases the accuracy estimates were mostly because of additive genetic relatedness between animals, rather than linkage disequilibrium between SNP and quantitative trait loci. Therefore, our results suggest that genomic prediction for nematode resistance and BW may be of value in closely related animals, but that with the current SNP chip genomic predictions are unlikely to work across breeds.     

Evaluation of different biological data and computational classification methods for use in protein interaction prediction

Protein–protein interactions play a key role in many biological systems. High‐throughput methods can directly detect the set of interacting proteins in yeast, but the results are often incomplete and exhibit high false‐positive and false‐negative rates. Recently, many different research groups independently suggested using supervised learning methods to integrate direct and indirect biological data sources for the protein interaction prediction task. However, the data sources, approaches, and implementations varied. Furthermore, the protein interaction prediction task itself can be subdivided into prediction of (1) physical interaction, (2) co‐complex relationship, and (3) pathway co‐membership. To investigate systematically the utility of different data sources and the way the data is encoded as features for predicting each of these types of protein interactions, we assembled a large set of biological features and varied their encoding for use in each of the three prediction tasks. Six different classifiers were used to assess the accuracy in predicting interactions, Random Forest (RF), RF similarity‐based k‐Nearest‐Neighbor, Naïve Bayes, Decision Tree, Logistic Regression, and Support Vector Machine. For all classifiers, the three prediction tasks had different success rates, and co‐complex prediction appears to be an easier task than the other two. Independently of prediction task, however, the RF classifier consistently ranked as one of the top two classifiers for all combinations of feature sets. Therefore, we used this classifier to study the importance of different biological datasets. First, we used the splitting function of the RF tree structure, the Gini index, to estimate feature importance. Second, we determined classification accuracy when only the top‐ranking features were used as an input in the classifier. We find that the importance of different features depends on the specific prediction task and the way they are encoded. Strikingly, gene expression is consistently the most important feature for all three prediction tasks, while the protein interactions identified using the yeast‐2‐hybrid system were not among the top‐ranking features under any condition. Proteins 2006. © 2006 Wiley‐Liss, Inc.     

The effect of optimal space allowance on growth performance and physiological responses of pigs at different stages of growth

This study was conducted to determine the optimal space allowance for maximizing the growth performance of pigs at each of the following five growth stages (based on BW ranges): stage 1, 11 to 25 kg BW; stage 2, 25 to 45 kg BW; stage 3, 45 to 65 kg BW; stage 4, 65 to 85 kg BW; and stage 5, 85 to 110 kg BW. A total of 1590 crossbred (Landrace×Yorkshire×Duroc) pigs were assigned to one of four treatments at each growth stage, with three replicates each. Pen areas at each growth stage were 6, 11, 16, 19.5 and 20 m² for stages 1 to 5, respectively. Space allowances for the four treatments at each growth stage were modified by varying the number of pigs per pen (22, 25, 28 and 31 pigs in T1, T2, T3 and T4, respectively). Blood samples were collected on the final day of each growth stage. The average daily gain (ADG) decreased significantly with decreased space allowances at all growth stages, except at stage 2. Average daily feed intake (ADFI) was not significantly affected by space allowances at stages 1 to 4; however, at stage 5, there was a linear effect of space allowance on ADFI. Thus, the feed conversion ratio showed results similar to those for ADG. Serum cortisol concentrations, indicating the level of stress response, increased as space allowances decreased. The highest serum cortisol concentrations were observed in T3 at stages 2 to 5. Serum tumor necrosis factor-α levels were significantly higher in association with a small space allowance than with at large space allowance at stages 2, 4 and 5. Serum interleukin-1β levels also increased in a significant linear manner at every growth stage in pigs reared at a low space allowance, except at stage 4 (P=0.068). This study found that limited space allowance decreases the growth performance of pigs and induces stress and inflammatory responses. We confirmed that no significant effect of space allowance on growth performance and serum cortisol concentrations are observed between T1 and T2 across all growth stages. We suggest that the optimal space allowances for pigs according to their BW are as follows: 0.24, 0.44, 0.64, 0.78 and 0.80 m²/pig for BWs of 11 to 25, 25 to 45, 45 to 65, 65 to 85 and 85 to 115 kg, respectively.     

A comparison of model selection methods for prediction in the presence of multiply imputed data

Many approaches for variable selection with multiply imputed data in the development of a prognostic model have been proposed. However, no method prevails as uniformly best. We conducted a simulation study with a binary outcome and a logistic regression model to compare two classes of variable selection methods in the presence of MI data: (I) Model selection on bootstrap data, using backward elimination based on AIC or lasso, and fit the final model based on the most frequently (e.g. ) selected variables over all MI and bootstrap data sets; (II) Model selection on original MI data, using lasso. The final model is obtained by (i) averaging estimates of variables that were selected in any MI data set or (ii) in 50% of the MI data; (iii) performing lasso on the stacked MI data, and (iv) as in (iii) but using individual weights as determined by the fraction of missingness. In all lasso models, we used both the optimal penalty and the 1‐se rule. We considered recalibrating models to correct for overshrinkage due to the suboptimal penalty by refitting the linear predictor or all individual variables. We applied the methods on a real dataset of 951 adult patients with tuberculous meningitis to predict mortality within nine months. Overall, applying lasso selection with the 1‐se penalty shows the best performance, both in approach I and II. Stacking MI data is an attractive approach because it does not require choosing a selection threshold when combining results from separate MI data sets     

Design of a low‐density SNP chip for the main Australian sheep breeds and its effect on imputation and genomic prediction accuracy

Genotyping sheep for genome‐wide SNPs at lower density and imputing to a higher density would enable cost‐effective implementation of genomic selection, provided imputation was accurate enough. Here, we describe the design of a low‐density (12k) SNP chip and evaluate the accuracy of imputation from the 12k SNP genotypes to 50k SNP genotypes in the major Australian sheep breeds. In addition, the impact of imperfect imputation on genomic predictions was evaluated by comparing the accuracy of genomic predictions for 15 novel meat traits including carcass and meat quality and omega fatty acid traits in sheep, from 12k SNP genotypes, imputed 50k SNP genotypes and real 50k SNP genotypes. The 12k chip design included 12 223 SNPs with a high minor allele frequency that were selected with intermarker spacing of 50–475 kb. SNPs for parentage and horned or polled tests also were represented. Chromosome ends were enriched with SNPs to reduce edge effects on imputation. The imputation performance of the 12k SNP chip was evaluated using 50k SNP genotypes of 4642 animals from six breeds in three different scenarios: (1) within breed, (2) single breed from multibreed reference and (3) multibreed from a single‐breed reference. The highest imputation accuracies were found with scenario 2, whereas scenario 3 was the worst, as expected. Using scenario 2, the average imputation accuracy in Border Leicester, Polled Dorset, Merino, White Suffolk and crosses was 0.95, 0.95, 0.92, 0.91 and 0.93 respectively. Imputation scenario 2 was used to impute 50k genotypes for 10 396 animals with novel meat trait phenotypes to compare genomic prediction accuracy using genomic best linear unbiased prediction (GBLUP) with real and imputed 50k genotypes. The weighted mean imputation accuracy achieved was 0.92. The average accuracy of genomic estimated breeding values (GEBVs) based on only 12k data was 0.08 across traits and breeds, but accuracies varied widely. The mean GBLUP accuracies with imputed 50k data more than doubled to 0.21. Accuracies of genomic prediction were very similar for imputed and real 50k genotypes. There was no apparent impact on accuracy of GEBVs as a result of using imputed rather than real 50k genotypes, provided imputation accuracy was >90%.     

Genetic constraints on wing pattern variation in Lycaeides butterflies: A case study on mapping complex,multifaceted traits in structured populations

Patterns of phenotypic variation within and among species can be shaped and constrained by trait genetic architecture. This is particularly true for complex traits, such as butterfly wing patterns, that consist of multiple elements. Understanding the genetics of complex trait variation across species boundaries is difficult, as it necessitates mapping in structured populations and can involve many loci with small or variable phenotypic effects. Here, we investigate the genetic architecture of complex wing pattern variation in Lycaeides butterflies as a case study of mapping multivariate traits in wild populations that include multiple nominal species or groups. We identify conserved modules of integrated wing pattern elements within populations and species. We show that trait covariances within modules have a genetic basis and thus represent genetic constraints that can channel evolution. Consistent with this, we find evidence that evolutionary changes in wing patterns among populations and species occur in the directions of genetic covariances within these groups. Thus, we show that genetic constraints affect patterns of biological diversity (wing pattern) in Lycaeides, and we provide an analytical template for similar work in other systems.     

Predicting the carcass chemical composition and describing its growth in live pigs of different sexes using computed tomographys

The aims of this study were (1) to evaluate the ability of computed tomography (CT) to predict the chemical composition of live pigs and carcasses, (2) to compare the chemical composition of four different sex types at a commercial slaughter weight and (3) to model and evaluate the chemical component growth of these sex types. A total of 92 pigs (24 entire males (EM), 24 surgically castrated males (CM), 20 immunocastrated males (IM) and 24 females (FE)) was used. A total of 48 pigs (12 per sex type) were scanned repeatedly in vivo using CT at 30, 70, 100 and 120 kg and slaughtered at the end of the experiment. The remaining 44 were CT scanned in vivo and slaughtered immediately: 12 pigs (4 EM, 4 CM and 4 FE) at 30 kg and 16 pigs each at 70 kg and 100 kg (4 per sex type). The left carcasses were CT scanned, and the right carcasses were minced and analysed for protein, fat, moisture, ash, Ca and P content. Prediction equations for the chemical composition were developed using Partial Least Square regression. Allometric growth equations for the chemical components were modelled. By using live animal and carcass CT images, accurate prediction equations were obtained for the fat (with a root mean square error of prediction (RMSEP_CV) of 1.31 and 1.34, respectively, and R²=0.91 for both cases) and moisture relative content (g/100 g) (RMSEP_CV=1.19 and 1.38 and R²=0.94 and 0.93, respectively) and were less accurate for the protein (RMSEP_CV=0.65 and 0.67 and R²=0.54 and 0.63, respectively) and mineral content (RMSEP_CV from 0.28 to 1.83 and R² from 0.09 to 0.62). Better equations were developed for the absolute amounts of protein, fat, moisture and ash (kg) (RMSEP_CV from 0.26 to 1.14 and R² from 0.91 to 0.99) as well as Ca and P (g) (RMSEP_CV=144 and 71, and R²=0.76 to 0.66, respectively). At 120 kg, CM had a higher fat and lower moisture content than EM. For protein, CM and IM had lower values than FE and EM. The ash content was higher in EM and IM than in FE and CM, while IM had a higher Ca and P content than the others. The castrated animals showed a higher allometric coefficient for fat and a lower one for moisture, with IM having intermediate values. However, for the Ca and P models, IM presented higher coefficients than EM and FE, and CM were intermediate.     

The welfare of growing pigs in five different production systems: assessment of feeding and housing

Ninety-one farms were visited over a 2-year period to assess the welfare of growing pigs in five different production systems found either in France or in Spain using the Welfare Quality® protocol. This study focused on animal-based measures as indicators of ‘good feeding’ and ‘good housing’. Multiple Generalized Linear Mixed Models were performed for each measure to evaluate the differences between production systems and to detect possible causal factors. Pigs in the conventional system presented the lowest prevalence of poor body condition, whereas extensive Mallorcan Black pigs and extensive Iberian pigs were associated with a decreased prevalence of bursitis and pig dirtiness. The straw-bedded system presented a lower prevalence of bursitis, but poorer hygiene and more susceptibility of poor body condition than the conventional system. The age of the animals had a significant effect on the appearance of bursitis in the three intensive systems studied. The type of floor was a significant causal factor of bursitis and pig dirtiness in the conventional system and among intensive Iberian pigs. The feeding system was another causal factor of pig dirtiness on more than 50% of the body in the conventional system, whereas pig dirtiness on less than 50% of the body was influenced by the age of the animals. The prevalence of huddling animals in the conventional system was associated with the highest stocking densities and the lowest environmental temperatures. The results indicate that there were important differences between production systems based on animal-based indicators of the good feeding and housing principles. The recording of the age of the animals, type of floor, feeding system, stocking density and environmental temperature can be useful to predict the appearance of a given welfare measure of ‘good housing’ on a farm.