- Threshold models are important statistical tools in animal breeding and quantitative genetics used to analyze traits that are recorded in categories rather than as continuous numerical measurements. Many economically important traits in livestock production, such as fertility, disease resistance, calving difficulty, survival, and litter size classifications, are expressed as discrete outcomes. These traits cannot always be analyzed effectively using traditional linear models because their observed categories may represent an underlying biological process that is continuous but cannot be measured directly. Threshold models address this challenge by assuming that an unobserved continuous variable, known as a latent liability, determines the category in which an animal’s observed performance falls. By connecting categorical observations to underlying genetic and environmental influences, threshold models help researchers estimate genetic parameters and breeding values more appropriately.
- In animal breeding, traits are commonly divided into continuous and categorical types. Continuous traits, such as body weight, milk yield, and body length, can be measured on a numerical scale and analyzed using linear models. Categorical traits, however, are recorded in distinct groups. For example, calving difficulty may be classified as unassisted, assisted, or requiring substantial intervention. Disease status may be recorded as affected or unaffected, while reproductive performance may be classified as pregnant or not pregnant. Although these categories are useful for recording and management, they may not fully represent the biological differences among animals. Two animals assigned to the same category may have different underlying genetic liabilities, even though their observed records are identical.
- The central idea behind a threshold model is that categorical outcomes arise when an underlying continuous liability crosses one or more thresholds. This liability cannot usually be observed directly, but it is influenced by genetic factors, environmental conditions, and random variation. For example, an animal’s liability to develop a disease may reflect its genetic susceptibility, nutrition, exposure to pathogens, housing conditions, and immune function. If the liability exceeds a particular threshold, the animal is classified as affected. If it remains below that threshold, the animal is classified as unaffected. The observed category therefore provides indirect information about the animal’s underlying liability.
- For a binary trait, a simplified threshold model can be written as:
- li=xi’+ai+ei
- Here, li represents the latent liability of animal i, xi’ represents the systematic effects of factors such as herd, year, season, age, and management group, ai represents the animal’s genetic effect, and ei represents the residual contribution of other influences. The observed outcome depends on whether the liability is above or below a defined threshold. In a binary model, the threshold can be set to zero for identification, so that the animal is classified into one category if its liability exceeds zero and the other if it does not. The numerical values of liability are not directly observed; the model estimates their distribution and the genetic effects that help explain the observed outcomes.
- The threshold concept is particularly useful because categorical traits often have a biological basis that varies continuously. Consider disease resistance in cattle. Animals may be recorded simply as diseased or healthy, but their underlying susceptibility can differ substantially. One animal may have a low liability to disease, while another may have a liability close to the threshold. Both may be classified as healthy under a particular set of conditions, but their genetic predispositions may differ. A threshold model uses the available information to estimate these differences in liability and can help predict which animals are more likely to produce resistant offspring.
- Threshold models are commonly used to analyze binary traits, which have two possible outcomes. Examples include whether a cow becomes pregnant following insemination, whether a calf survives to a specified age, whether an animal develops a particular disease, or whether a ewe successfully rears a lamb. These traits are important because they directly affect the efficiency and profitability of livestock production. However, their genetic evaluation can be challenging because the observed records contain only limited information about the underlying biological variation. Threshold models provide a statistical framework for analyzing these outcomes while accounting for relationships among animals and differences in environmental conditions.
- They can also be used for ordinal categorical traits, in which observations are arranged in a meaningful order. Examples include calving ease scores, disease severity classifications, temperament scores, and body condition categories. In these cases, more than one threshold separates the possible outcomes. For instance, calving difficulty may be classified into several ordered categories ranging from an uncomplicated birth to increasingly difficult deliveries. Each category corresponds to a range of latent liability values. Animals with liabilities below the first threshold fall into the lowest category, while animals with higher liabilities fall into progressively more severe categories. This approach allows the model to use the ordering of the categories without assuming that the numerical differences between category labels are equal.
- An important application of threshold models is the genetic evaluation of fertility traits. Reproductive performance is often recorded as a categorical outcome, such as conception success or failure, pregnancy status, or whether an animal produces offspring within a defined period. These traits can have major economic consequences because poor fertility increases the cost of maintaining breeding animals and reduces the number of offspring produced. Threshold models can account for genetic relationships among animals and help estimate differences in reproductive liability. The resulting breeding values can support selection programs aimed at improving fertility, provided that reproductive records are sufficiently reliable and the model accounts for relevant management and environmental effects.
- Threshold models are also useful in the evaluation of calving ease. Difficult births can lead to calf mortality, injury, reduced milk production, delayed reproductive recovery, and increased veterinary costs. Calving ease is often recorded in ordered categories because the degree of assistance required varies from one birth to another. A threshold model can relate these categories to the underlying liability associated with calving difficulty. Genetic effects for the calf and the dam may both be relevant, depending on the evaluation objective. Direct genetic effects can influence characteristics such as calf size and conformation, while maternal genetic effects may influence the dam’s ability to give birth successfully. Accounting for these effects can improve breeding decisions aimed at reducing difficult calvings while maintaining desirable growth and production traits.
- Another important application is disease resistance. Livestock diseases reduce productivity, increase treatment costs, and may affect animal welfare and survival. Many disease traits are recorded as affected or unaffected, even though disease development depends on a combination of genetic susceptibility, pathogen exposure, management, and environmental conditions. Threshold models can help separate genetic variation from systematic environmental differences when the data and model are appropriate. This supports the identification of animals with a lower genetic liability to disease and can contribute to breeding programs that improve health and reduce reliance on treatment. Nevertheless, accurate disease recording and appropriate control for differences in exposure are essential, because an animal recorded as healthy may simply have had less opportunity to encounter a pathogen.
- Survival and mortality traits are additional areas where threshold approaches may be useful. For example, calf survival to a particular age can be recorded as a binary outcome, while lamb survival and piglet survival are important measures in sheep and pig production. These outcomes are affected by genetics, maternal performance, birth weight, disease, weather, housing, nutrition, and management practices. A threshold model can be used to estimate genetic liability associated with survival, although traits involving the timing of death or the length of time an animal remains alive may be better analyzed using survival analysis models. The appropriate method depends on whether the main objective is to predict a categorical outcome or to model the time until an event occurs.
- Threshold models are closely related to generalized linear mixed models, particularly logistic and probit models. These approaches are designed for outcomes that do not follow a normal distribution, such as binary or categorical records. A logistic model uses a logistic link function, while a probit threshold model uses the cumulative standard normal distribution to relate the linear predictor to the probability of an observed outcome. In quantitative genetics, threshold models often use a latent liability formulation and a probit link. Although logistic and probit models differ in their mathematical assumptions and scaling, both can be used to model categorical responses while incorporating genetic and environmental effects.
- The choice between a threshold model and a conventional linear model depends on the nature of the trait, the structure of the data, and the purpose of the genetic evaluation. Linear models are often appropriate for continuous traits that are reasonably described by their assumed residual distribution. Applying a linear model directly to a binary trait can produce limitations because predicted values may fall outside the range of zero to one, and the residual variance of a binary outcome depends on its mean. Threshold models are designed to represent categorical outcomes more naturally. However, they are not automatically superior in every situation. Linear models may remain useful for some practical evaluations, particularly when categories are numerous or when operational simplicity and established evaluation systems are important. Model selection should consider statistical performance, data availability, interpretability, and the breeding objective.
- One of the key advantages of threshold models is their ability to incorporate pedigree relationships and genomic information. Animals are genetically related through shared ancestors, and these relationships provide information about their expected genetic similarity. By including a genetic covariance structure based on pedigree or genomic relationships, a threshold model can use records from relatives to improve the prediction of an animal’s genetic merit. This is particularly valuable for traits with low heritability or limited individual records. For example, a bull may have no direct record of calving ease in the sense of personally giving birth, but information from its offspring and relatives can help estimate its genetic contribution to calving difficulty. Genomic information may further improve prediction when suitable reference populations and validated evaluation methods are available.
- Threshold models can be used to estimate genetic parameters, including heritability on the underlying liability scale. In quantitative genetics, heritability describes the proportion of variation attributable to genetic differences, but categorical traits require special care in interpretation. A threshold model commonly estimates genetic variance and residual variance on a latent scale, because the liability itself is not directly observed. Heritability on this scale is not identical to the proportion of variation in the observed categories that is explained by genetics. Researchers may convert liability-scale estimates to an observed scale when needed, using assumptions about the distribution of liability and the frequency of the categories. Such conversions depend on the population’s incidence rate and should be interpreted carefully.
- The incidence of a categorical trait can strongly affect the amount of information available for genetic evaluation. For a binary trait, incidence refers to the proportion of observations in one category, such as the proportion of animals affected by a disease or the proportion of cows that become pregnant. If almost all animals fall into one category, there may be relatively little information for estimating genetic differences, particularly when the less frequent outcome is the one of interest. Rare outcomes can also create computational and statistical challenges. Large datasets, reliable records, and appropriate model specification can help, but they do not eliminate the need to assess whether the data provide sufficient information for accurate genetic evaluation.
- The estimation of threshold models is often more computationally demanding than fitting a simple linear model. The observed data provide categories rather than direct measurements of liability, so the model must estimate genetic effects and other parameters while accounting for the unobserved underlying distribution. Estimation methods may include Bayesian approaches, likelihood-based procedures, and specialized algorithms for generalized mixed models. Bayesian methods are frequently used for complex threshold models because they can accommodate hierarchical structures and complex prior distributions. Depending on the model and dataset, methods such as Gibbs sampling or other Markov chain Monte Carlo procedures may be used. The choice of method affects computational requirements, convergence assessment, and the uncertainty associated with parameter estimates.
- As with other genetic evaluation models, threshold models require appropriate fixed and random effects. Factors such as herd, year, season, age, parity, management group, and environmental conditions may influence the observed category. If important systematic effects are omitted, the model may attribute environmental differences to genetic variation. For example, differences in disease incidence between herds may result from vaccination programs, housing, pathogen exposure, or recording practices rather than differences in genetic resistance. Proper data editing and the inclusion of relevant explanatory variables are therefore essential for producing useful breeding values.
- Another challenge is the quality and consistency of categorical records. If different farms use different scoring systems or apply category definitions inconsistently, the observed records may not be directly comparable. For example, one herd may classify moderate calving difficulty as assisted birth, while another may use the same category only for more severe cases. Such differences can introduce bias and reduce the accuracy of genetic comparisons. Standardized definitions, clear recording protocols, and appropriate adjustments for systematic differences are important for successful threshold-model applications. Researchers must also consider whether the categories genuinely represent an ordered scale and whether the model’s assumptions are appropriate.
- Threshold models are particularly relevant to modern breeding programs because selection objectives increasingly include health, welfare, fertility, survival, and resilience alongside production. Historically, many livestock breeding programs emphasized traits such as growth rate, milk yield, or carcass composition because they were relatively easy to measure on a continuous scale. However, improving productivity alone does not guarantee better overall animal performance. Poor fertility, high disease incidence, difficult births, and early mortality can reduce the economic and biological efficiency of a production system. Threshold models provide tools for incorporating such categorical traits into genetic evaluation and balanced selection indexes.
- In genomic selection, threshold models can be combined with genomic relationships to predict the genetic liability of animals for categorical traits. Genomic prediction is especially attractive for traits that are expensive, difficult, or impossible to measure directly in breeding candidates. For example, genetic liability to a disease may be predicted using genomic data and disease records from related animals. This can help breeders select young animals before they have their own performance records. However, prediction accuracy depends on the size and relevance of the reference population, the quality of phenotypic records, genetic relationships between reference and candidate animals, and the stability of genetic effects across populations and environments.
- Threshold models can also be incorporated into multi-trait evaluations. A categorical trait may be genetically correlated with one or more continuous traits. For example, calving difficulty may be associated with birth weight, and disease resistance may be associated with production or other health traits. A multi-trait model can use these genetic relationships to improve prediction and help breeders understand possible trade-offs. Such models can be valuable when direct records for a categorical trait are limited but correlated traits are more widely recorded. Their usefulness depends on the accuracy of the estimated genetic correlations and on the biological relevance of the traits included.
- In conclusion, threshold models are valuable statistical tools in animal breeding and quantitative genetics because they analyze categorical traits through an underlying continuous liability. They are widely relevant to fertility, calving ease, disease resistance, survival, and other traits that cannot be measured adequately on a continuous scale. By incorporating genetic relationships, environmental effects, pedigree information, and genomic data, threshold models can support the estimation of genetic parameters and breeding values for important functional traits. Although these models require careful data recording, appropriate assumptions, and potentially intensive computation, they help breeders evaluate a broader range of characteristics and develop livestock populations that are productive, healthy, fertile, and better suited to sustainable animal production.
