Quantitative genetics studies how inherited differences and environmental conditions contribute to variation in measurable traits, and how that variation relates to resemblance, prediction and change across generations. Its gold standard is a model whose population, measurement, assumptions and uncertainty remain visible. A variance estimate is useful when we know what was measured, which comparisons identify it and where the result can reasonably transfer.
This guide connects trait measurement, genetic variance, family resemblance, selection, correlated traits and genomic prediction. Original fictional examples show the calculations and their limits. They are learning exercises, not estimates for a real species or instructions for making medical, educational or breeding decisions about individuals.
Choose your quantitative genetics route
Connect trait variation, family resemblance, selection, covariance and validated prediction.
Route 1 · Chapters 1–3: Define traits and variation
Route 2 · Chapters 4–6: Understand inheritance and resemblance
Route 3 · Chapters 7–9: Predict selection and environmental responses
Route 4 · Chapters 10–11: Build and validate genomic predictions
Route 5 · Chapters 12–14: Transfer, diagnose and continue
Chapter index
Define traits and variation · Chapters 1–3
Understand inheritance and resemblance · Chapters 4–6
Predict selection and environmental responses · Chapters 7–9
Build and validate genomic predictions · Chapters 10–11
Transfer, diagnose and continue · Chapters 12–14
Consider a collection of plants whose flowering dates differ. One question asks whether relatives tend to flower at similar times. Another asks whether a particular DNA variant changes flowering through a known molecular pathway. A third asks how accurately a model predicts flowering in a new season. These questions are connected, but success at one does not automatically answer the others.
Quantitative genetics focuses on variation and its statistical structure. It can study continuously measured characteristics, counts and traits analysed through other suitable statistical models. The biological feature must be defined before choosing the model. Flowering date, number of flowers and probability of flowering are different outcomes, even when collected from the same plants.
The NHGRI definition of a polygenic trait describes a characteristic influenced by multiple genes. Environmental influences can also contribute. A quantitative measurement does not by itself prove that thousands of genes are involved; the observed distribution cannot uniquely reveal the underlying genetic architecture.
Original measurement brief
Suppose our fictional study records days from germination to the first open flower, using the same observation rule for every plant. The population is a specified set of related plants grown during one season. Plants that never flower before observation ends need a stated treatment in the analysis. Assigning them the final observation day as though they flowered then would change the outcome being measured.
Write the question in operational terms: how much of the observed variation is associated with additive inherited differences under the model, and how well does that information predict a later cohort? This wording leaves room for uncertainty and separates explanation of existing records from prediction of new records.
The Gold Standard of Genetics supplies the underlying language of genes, alleles and inheritance. The quantitative branch adds a particular discipline: every claim about variation needs a defined distribution, not merely a familiar trait name.
A mean describes the centre of a set of measurements. A variance describes squared deviations around that centre. Two populations can have the same mean and different variances, or different means and the same variance. Neither comparison alone tells us which genetic or environmental process produced the pattern.
Original four-value calculation
Take fictional measurements of 8, 10, 12 and 14 centimetres. Their mean is 11 cm. Deviations from the mean are −3, −1, 1 and 3 cm, and the squared deviations sum to 20 cm². If these four values constitute the complete population being described, variance is 20/4 = 5 cm². The corresponding standard deviation is the square root of five, approximately 2.236 cm.
If instead the four values are treated as a sample for estimating a wider population variance, the usual unbiased sample variance uses denominator three, giving approximately 6.667 cm². The different answers reflect different statistical tasks. A report must state which one it uses rather than choosing a denominator because it produces a preferred result.
Now add five centimetres to every measurement. The mean becomes 16 cm, but all deviations from the mean remain unchanged. Variance stays at 5 cm² under the complete-population convention. A common environmental improvement could therefore shift the mean without necessarily changing the variance in this simplified example.
Multiply every measurement by ten to express millimetres. The mean multiplies by ten, variance by one hundred and standard deviation by ten. This provides a useful unit check. A variance ratio can remain unchanged under a simple unit conversion even though its numerator and denominator become much larger numbers.
These elementary distinctions prevent later mistakes. A variance component measured in squared units cannot be added directly to a mean measured in ordinary units. Nor does a larger population mean imply more genetic variation. Start with the descriptive quantities before asking what biological model might explain them.
A simple starting model writes an observed phenotype as a population mean plus a genetic contribution and a residual contribution. The residual may include environmental effects, developmental variation and measurement error that the model has not separately represented. It should not automatically be interpreted as a single identifiable environmental cause.
If the genetic and residual components are uncorrelated in the specified model, their variances add. If they are correlated, the variance of their sum also includes twice their covariance. Omitting that term is an assumption, not an algebraic identity that holds for every biological dataset.
Original covariance check
Suppose two centred model components have variances 9 and 16 units squared. With zero covariance, total variance is 25. With covariance 3, total variance becomes 9 + 16 + 2 × 3 = 31. With covariance −3, it becomes 19. All three calculations use the same component variances but different relationships between the components.
The example does not say we can directly observe a genetic contribution and a residual contribution for each individual. Usually those are inferred through a design and model. The calculation shows why a decomposition requires more than a list of plausible causes.
A study might represent block, season, maternal environment or permanent individual effects separately. Doing so changes what remains in the residual. Comparing residual variances from two differently specified models without noticing this can create an apparent scientific disagreement that is partly a difference in bookkeeping.
Inspect how each component is identified. Does the dataset contain relatives across several environments, or is every family confined to one location? Are repeated records available? Is there enough independent information to distinguish family resemblance from a shared growing environment? More elaborate software cannot create contrasts absent from the study.
The practical test is explanatory: state which observations would change if a particular component were larger. If no available comparison distinguishes two components, acknowledge the limitation rather than treating the fitted labels as proof that the causes have been separated.
CHAPTER 4 OF 14
4. Understand additive effects without denying biological interactions
Additive genetic variance concerns the part of genetic variation represented by the summed average effects of alleles in a specified population. Dominance concerns departures associated with combinations of alleles at a locus; epistasis concerns interactions involving loci. These statistical descriptions should not be confused with a claim that molecular processes operate independently.
Hill, Goddard and Visscher’s research on additive genetic variance demonstrates why substantial additive variance can coexist with non-additive gene action. The population distribution of alleles matters to the statistical partition. Detecting useful additive prediction therefore does not establish that biological interactions are absent.
Original genotype illustration
Assign a hypothetical locus genotype means of 10, 14 and 14 units for aa, Aa and AA. The heterozygote lies above the midpoint of the two homozygotes: that midpoint is 12. The departure illustrates dominance on this measurement scale. It does not establish the frequencies of the genotypes or the variance they generate in a population.
If nearly every organism has the same genotype, the locus contributes little observed variation despite the differences between the hypothetical genotype means. If the genotypes are all common, the same genotype values can produce more variation. An effect and a variance are related but different quantities.
The average effect used in prediction depends on the population in which it is defined. Changes in allele frequencies and associations can change that statistical effect. A coefficient should therefore travel with its reference population and modelling assumptions rather than being presented as a timeless property of a DNA letter.
For learning, draw three separate boxes on paper: the biological interaction, the genotype means under particular conditions, and the distribution of genotypes in the population. Ask which evidence belongs in each box. This prevents a mechanistic observation from being stretched into an unsupported population variance estimate, or a statistical variance estimate from being mistaken for a complete molecular explanation.
Narrow-sense heritability is conventionally written h² = VA/VP, with VA the additive genetic variance and VP the phenotypic variance under the specified model. Broad-sense heritability includes the genetic variance represented by the broader genetic component. Which quantity is appropriate depends on the question and the design; the names are not interchangeable.
Original two-environment calculation
Assume additive genetic variance is 12 units squared and all other modelled variance is 18, with zero covariance. Phenotypic variance is 30 and h² = 12/30 = 0.40. Now consider a second environment in which additive variance remains 12 but the other variance falls to 8. The ratio becomes 12/20 = 0.60.
The second ratio is larger although the stipulated additive variance is unchanged. A change in the denominator is enough. We must not conclude that genes became more powerful or that an individual is now sixty per cent genetically determined. The statement concerns differences among observations under a particular model.
Conversely, a trait can respond strongly to an environmental intervention even when its variation within the original environment has high heritability. A ratio describing existing variation does not bound the change possible under a new condition. Moving a whole distribution and explaining differences within it are separate tasks.
The existing specialist Heritability, Selection Response and the Environment Behind the Estimate develops these distinctions and their selection consequences in detail. Use it when the immediate task is interpreting or calculating a heritability estimate; the present guide places that estimate among the field’s other tools.
In a report, include the trait definition, population, environment, observational unit, estimation method and uncertainty. An estimate for individual records differs from one for means over repeated observations. A number copied without that context may be precise in appearance while answering the wrong question for its new reader.
Relatives share inherited material, but they may also share environments and developmental influences. A resemblance between relatives is informative only in relation to those alternatives. The statistical model must explain which patterns of resemblance are attributed to inheritance and which can arise through shared conditions.
Original family-location problem
Imagine ten fictional plant families. Each family is grown in its own tray, and all members of a family receive the same tray conditions. Family one grows taller than family two. A family effect and a tray effect are now entangled: the observations do not independently vary family and tray.
Increasing the number of plants in each tray gives more precise information about each combined family–tray outcome. It does not by itself separate the two explanations. A thousand plants arranged in the same confounded design can preserve the same inferential problem as a hundred.
A more informative design distributes members of each family across several trays and places several families within each relevant environmental block. Randomisation and replication create comparisons that the original layout lacked. The exact design must fit the organisms and question; this is a reasoning example rather than a universal experimental recipe.
Parent–offspring regression has related assumptions. In an appropriate simple additive setting, offspring values regressed on mid-parent values can inform narrow-sense heritability. Shared environment, assortative mating, maternal effects and selection of the sampled families can complicate that interpretation. The slope is first a fitted relationship; its genetic meaning requires justification.
Consider a fictional slope of 0.4. It predicts a four-unit difference in mean offspring phenotype for a ten-unit difference in the chosen parental predictor, within the fitted model. It does not say every offspring receives forty per cent of its measured trait from its parents, nor does it describe a fraction of DNA transmitted.
For any resemblance estimate, ask which comparison would reveal a shared-environment explanation. That question often identifies the most useful design improvement before another statistical method is considered.
A selection differential compares the mean phenotype of selected parents with the original population mean. A response compares an appropriate descendant mean with its baseline. These quantities belong to different stages of the process. Their relationship is a modelled expectation, not a promise that every descendant resembles the selected parents exactly.
Original one-generation calculation
Suppose a fictional population has mean seed mass 40 mg, and selected parents average 45 mg. The selection differential S is 5 mg. Under a simple additive model with h² = 0.30 and suitable assumptions, the breeder’s equation predicts R = h²S = 1.5 mg. If the baseline conditions remain comparable, the predicted descendant mean is 41.5 mg.
Now suppose selected descendants average 44 mg, but an unselected control grown alongside them averages 42.5 mg. Their contrast is still 1.5 mg. Comparing the selected descendants only with the original 40 mg mean would attribute the entire four-milligram difference to selection and overlook the changed baseline.
The control comparison is itself useful only if it is appropriate. Differential growing conditions, changes in sampling or different measurement procedures can undermine it. A control is a part of an argument, not a word that automatically repairs every confounder.
Morrissey, Kruuk and Wilson’s analysis explains difficulties in applying the breeder’s equation to observational natural populations. An association between a phenotype and reproductive success need not identify the genetic covariance responsible for evolutionary change. Correlated traits and environmental causes of both phenotype and success can matter.
Do not multiply a one-generation response by ten and call it a ten-generation forecast without further justification. Genetic variance, allele frequencies, environmental conditions and selection can change. A short prediction can be useful while a long extrapolation from the same coefficient is unsupported.
The linked heritability specialist provides further worked comparisons. The wider quantitative-genetic habit is to preserve the chain from measured differential through model assumptions to a separately checked response.
Traits often vary together. A phenotypic covariance describes their observed co-variation; an additive genetic covariance concerns co-variation between their additive genetic values under the model. The two need not be equal. Shared environment can create an observed association without the same inherited relationship.
The multivariate framework follows Lande’s quantitative analysis of multivariate evolution: a genetic covariance matrix links a vector of selection gradients to predicted changes in trait means under the model.
Original two-trait calculation
Use two fictional standardised traits, with additive variance one for each and additive covariance −0.4. The genetic covariance matrix therefore has diagonal entries 1 and off-diagonal entries −0.4. Suppose the supplied selection-gradient vector is 0.20 for the first trait and zero for the second. Under the multivariate linear approximation, multiplying the matrix by the gradient gives predicted changes of 0.20 and −0.08.
The second trait changes despite having a zero direct gradient in this model. Its genetic covariance with the first trait transmits a correlated response. If the covariance were positive 0.4 instead, its predicted response would be positive 0.08. The sign of the relationship matters to the practical interpretation.
Now supply gradients 0.20 and 0.10 with the original negative covariance. The predicted first response becomes 1 × 0.20 − 0.4 × 0.10 = 0.16. The second becomes −0.4 × 0.20 + 1 × 0.10 = 0.02. Both gradients are positive, but the responses are constrained by the covariance structure in this example.
These are calculations from supplied quantities, not instructions for estimating gradients or proving causal selection. Trait scaling matters: rescaling a trait changes the relevant numerical entries and coefficients. A matrix copied between incompatible units can produce a formally neat but biologically meaningless answer.
When several outcomes matter, define them before choosing an index or target. Increasing one outcome while ignoring a correlated cost can make a narrow result look better than the overall outcome. Life History Theory provides a complementary route into survival, growth and reproductive allocation; quantitative genetics asks how inherited covariation affects change among those components.
Genetic differences need not have identical consequences in every environment. A genotype-by-environment interaction occurs when the difference between genetic backgrounds changes with environment on the chosen measurement scale. It can change the magnitude of a difference, its direction, or both.
Original reaction-norm case
Suppose fictional line A yields 12 units under condition one and 8 under condition two. Line B yields 10 and 11 respectively. A exceeds B by two units in the first condition but falls below B by three in the second. The ranking reverses. Reporting only that A is the better line would discard the environmental condition that makes the statement true.
If the target setting consists of eighty per cent condition one and twenty per cent condition two, the weighted means are 0.8 × 12 + 0.2 × 8 = 11.2 for A and 0.8 × 10 + 0.2 × 11 = 10.2 for B. If the proportions reverse, the means become 8.8 and 10.8. The preferred prediction changes with the target environment distribution.
These averages assume the supplied outcomes and weights are appropriate. They do not include uncertainty, interactions among organisms or a cost attached to variability itself. If a decision is especially sensitive to a poor season, comparing means alone may omit the outcome that matters most.
The specialist phenotypic plasticity guide develops reaction norms and distinguishes environmental responses from inherited population change. For quantitative genetics, the additional question is how those responses affect variance, covariance and prediction in the target setting.
Check measurement scale before interpreting an interaction. Absolute differences and proportional differences can tell different stories. Also avoid treating two environments as a complete map of all possible conditions. An observed crossing establishes a limitation of a universal ranking; it does not tell us the shape of the response outside the tested range.
A genetic association study asks whether variants and a phenotype are statistically related under an analysis. A prediction model asks how well available information forecasts an outcome in new observations. A variant can have a strong mechanistic role yet add little predictive information in a particular sample; a useful predictor can also tag another causal process rather than reveal it directly.
Yang and colleagues’ 2010 study of human height helped demonstrate the value of estimating aggregate contributions from many common markers rather than considering only individually significant associations. The methodological distinction matters: variation captured jointly by measured markers is not the same quantity as variance attributed to a list of discovered significant loci.
Meuwissen, Hayes and Goddard’s 2001 work investigated prediction using dense genome-wide markers. This provides a foundational route into genomic prediction, where information across markers is combined rather than requiring each marker to pass a separate discovery threshold before it can contribute.
Original score calculation
Imagine a teaching score with three supplied weights: 0.2, −0.1 and 0.4. An individual has coded allele counts 2, 1 and 0 at the corresponding markers. The weighted sum is 2 × 0.2 + 1 × (−0.1) + 0 × 0.4 = 0.3. This arithmetic gives a score on the model’s scale. Without the intercept, outcome definition and calibration, it is not a predicted phenotype or probability.
Marker coding must match the weights. Reversing which allele is counted changes the calculation. Correlated markers also require appropriate handling in model construction; treating several linked signals as independent can distort the score. These are reasons to retain the model specification, not merely its final output column.
An association does not establish a molecular mechanism, and a predictive score does not describe destiny. The NHGRI explanation of polygenic risk scores emphasises probabilistic interpretation. Clinical usefulness requires additional evidence beyond a statistically associated or well-ranked score.
A model can fit its training data closely and still predict new observations poorly. Evaluation needs data that were not used to choose coefficients, tune settings or select a favourable model. The split should also represent the intended transfer: new relatives, new families, a future season and a different population are different tests.
Original error comparison
Suppose four held-out observations are 10, 12, 14 and 16 units. Predictions are 11, 11, 13 and 17. Errors are 1, −1, −1 and 1, giving mean squared error one and root mean squared error one unit. The mean error is zero, but that does not mean the predictions are exact; positive and negative errors cancel in the average.
Now add three to every prediction. The errors become 4, 2, 2 and 4. The ranking of individuals is unchanged, but the mean error is three and mean squared error is ten. This illustrates why ranking ability and calibration are different. A model can order observations reasonably while systematically overpredicting their level.
If closely related individuals appear in both training and test sets, evaluation may answer a within-family prediction question. It does not necessarily establish performance for unfamiliar families. Group-aware splits can address the latter question, although the appropriate grouping depends on the intended application and available data.
Data preparation can also leak information. Calculating an outcome-informed transformation or selecting markers using the full dataset before splitting lets test information influence training. Perform the relevant fitting steps inside the training process, then apply the resulting procedure to held-out records.
Report the target population, sample size, error measures and uncertainty alongside performance. Examine subgroups and environmental conditions when they are relevant to use. A single average can conceal a failure concentrated in precisely the setting where the prediction is needed. Successful validation is evidence for a defined transfer, not a permanent certificate for every future setting.
The most revealing test changes the background assumptions while preserving the familiar vocabulary. Suppose an original fictional study has additive variance 10, other variance 15 and zero covariance. Its narrow-sense heritability is 10/25 = 0.40. Selected parents exceed the original mean by six units, giving a simple predicted response of 2.4 units.
Original transfer question and answer
A new study uses means of four repeated measurements per individual. Assume the other variance consists entirely of independent measurement noise of variance 15 per record, while the additive component is stable across records. Averaging reduces that noise variance to 15/4 = 3.75. The variance ratio for the mean record is now 10/13.75, approximately 0.727.
The increase reflects a changed observational unit and reduced noise, not a sudden change in inheritance. It also does not justify multiplying the original six-unit selection differential by 0.727 unless that differential was measured on the corresponding predictor and the selection model is appropriate. Quantities from different definitions cannot be combined merely because they use the same physical units.
Now change the assumption again. Suppose the fifteen residual units comprise permanent individual variance nine plus independent noise six. Averaging four records leaves nine unchanged and reduces six to 1.5. The resulting ratio is 10/(10 + 9 + 1.5), approximately 0.488. Repeated measurement cannot average away a stable individual difference in the same way as independent noise.
This exercise identifies the mechanism behind a common overclaim. “More measurements improve the estimate” is incomplete: which component becomes better estimated, and which source of variation is reduced by averaging? Repeating a biased measurement or repeatedly observing the same unrepresentative conditions may preserve the relevant problem.
To transfer the reasoning correctly, rewrite the phenotype definition, covariance assumptions and selection criterion. Then recalculate. The answer can change substantially while every arithmetic operation remains elementary. The difficulty lies in preserving the biological and statistical meaning of the numbers as the study changes.
Does high heritability mean the environment is unimportant?
No. Heritability concerns variation in a specified setting. It does not measure the total causal importance of an environment to development, nor does it predict every response to an intervention. A population can have limited environmental variation during a study while responding strongly when conditions are changed beyond that range.
Can a family study prove which genes cause a trait?
It can provide evidence about patterns of inherited variation under a model, but it does not necessarily identify particular causal variants or molecular mechanisms. Mapping, functional investigation and appropriate causal designs answer additional questions. Keep the level of the claim aligned with the level of the evidence.
Does additive variance mean genes act independently?
No. Additivity in a statistical representation is different from independence of molecular processes. Population allele frequencies and the structure of the model affect the decomposition. A useful additive predictor can coexist with dominance, interactions and complex development.
Why do estimates differ between studies?
Check the trait definition, age, population, environment, measurement quality, relatedness information and model before declaring a contradiction. Different samples can genuinely contain different genetic and residual variances. Sampling uncertainty also matters. A difference between point estimates is not by itself evidence of a meaningful difference between the underlying parameters.
Can within-population estimates explain differences between populations?
Not automatically. Explaining variation within one setting and explaining a difference in means across settings require different comparisons. Environmental, historical and sampling differences may contribute to the latter. A within-population variance ratio does not partition a between-population mean difference into genetic and environmental percentages.
What should make me distrust a polished result?
Look for missing outcome definitions, relatives treated as independent without justification, family and environment confounded, uncertainty omitted, or prediction assessed on data used for model choice. The repair should target the missing comparison or assumption. Adding more decimal places does not supply missing evidence.
Begin with a small invented dataset and calculate a mean, variance and covariance by hand. State whether it is a complete population description or a sample estimate. Convert units and check that means and standard deviations change linearly while variances change by the square of the conversion factor. These checks develop habits that remain useful in much larger analyses.
Next, draw a family-by-environment layout. Decide which comparisons identify a family effect and which identify an environmental effect. Deliberately create a confounded layout, then repair it on paper. This teaches why experimental design and sampling are part of quantitative reasoning rather than administrative details completed before the real analysis.
Then work a one-generation response model and a two-trait covariance model. Change one assumption at a time. Ask whether the conclusion changes because the supplied biology changed, the observational unit changed or the arithmetic was wrong. Those are different diagnoses and call for different repairs.
For research reading, identify the estimated quantity before reading the headline result. Is it pedigree-based additive variance, marker-captured variance, a prediction accuracy or an association coefficient? Record the population and validation setting alongside the number. Do not compare estimates with different definitions as though they measured an identical object.
For the connected subject route, use Genetics for inheritance and Population Genetics for allele-frequency processes. Follow the heritability specialist for selection calculations and the plasticity specialist for environmental responses. Life History Theory connects inherited variation with survival and reproductive schedules.
The Gold Standard of Education explains the value of testing knowledge in changed situations. How X Works and The Gold Standard master provide broader reading routes. A strong final explanation should let another reader reconstruct the question, calculation, assumptions and limit of the conclusion without guessing what any important number means.