Twin Studies and Heritability
Identical twins raised in the same household sometimes end up strikingly different. This applies to personality, disease risk, or body weight. Meanwhile, fraternal twins can end up eerily alike. Why should genetically identical people ever differ? Why should less-related siblings ever converge? The twin study was designed to answer a narrower, more tractable version of this puzzle. For a given trait in a given population, how much of the variation among people traces to genetic differences rather than environmental ones? This is the question of Heritability: Broad and Narrow Sense. Twins offer a natural, if imperfect, experiment for estimating it. The logic is deceptively simple. Monozygotic (MZ) twins share essentially 100% of their segregating genetic variants. They arose from a single fertilized egg that split in two. Dizygotic (DZ) twins are ordinary siblings who happen to share a womb and a birth date. They share on average 50% of their segregating alleles, just like any full siblings. If MZ twins resemble each other more than DZ twins do, on a trait, the excess resemblance is attributed to the extra genetic sharing. From that excess, a heritability estimate can be computed. This method was formalized largely through the work of twin registries and behavioral geneticists across the twentieth century. It has been used to estimate heritability for thousands of traits, from height and schizophrenia to educational attainment and divorce risk. Twin studies matter because they are one of the few tools available. They let researchers partition phenotypic variance into genetic and environmental components. This is done without controlled breeding experiments, which are ethically impossible in humans. They underlie major claims in psychiatric genetics, behavioral genetics, and medicine. They were the primary evidence base for heritability estimates before the genomic era. But the method's simplicity conceals a stack of assumptions. The "equal environments assumption" is chief among them. When violated, these assumptions can inflate or bias heritability estimates. This has generated real scientific controversy. This article works through the biology of twinning itself. It covers the classical twin design (ACE model) and how concordance and correlation data are converted into heritability estimates. It also covers the historical case that established the method and the disturbing history entangled with it. Two worked numerical examples follow. Finally, it addresses the major pitfalls that any careful reader of twin-study heritability numbers needs to know.
- MZ twins share ~100% of segregating alleles (same zygote, split embryo); DZ twins share ~50% on average, like ordinary siblings.
- Twin heritability estimates compare trait resemblance (concordance or correlation) in MZ vs. DZ pairs; greater MZ resemblance implies a genetic contribution.
- The classical "ACE model" partitions phenotypic variance into Additive genetic (A), shared/Common environment (C), and unique/non-shared Environment (E) components.
- Falconer's formula, h² ≈ 2(r_MZ − r_DZ), gives a quick narrow-sense heritability estimate from twin correlations.
- The key assumption is the "equal environments assumption" (EEA): MZ and DZ twins experience equally correlated environments for the trait in question — often violated, and much debated.
- Heritability from twin studies is population- and time-specific, says nothing about individuals, and does not partition genetic vs. environmental causes of group differences.
Twin studies estimate heritability by asking whether pairs who share (nearly) all their genes resemble each other more, for a trait, than pairs who share only half — with the excess MZ resemblance attributed to genetic variance, contingent on the assumption that both twin types share their environments equally.
The core idea
A twin study exploits a naturally occurring difference in genetic relatedness between two classes of sibling pairs. Monozygotic (MZ, "identical") twins arise when a single fertilized zygote splits into two embryos early in development; barring rare post-twinning mutations, they are genetically identical at essentially every locus in the genome. Dizygotic (DZ, "fraternal") twins arise from two separate eggs fertilized by two separate sperm in the same ovulatory cycle; genetically they are ordinary full siblings who happen to be the same age, sharing on average 50% of their segregating (variable) alleles, exactly as any two full siblings would. Both twin types, however, typically share the same womb, the same birth cohort, often the same household, school, and social environment. This gives the classical twin design its logical structure: MZ and DZ pairs are assumed to be exposed to equally similar environments (the "equal environments assumption," EEA), but MZ pairs share twice the genetic relatedness of DZ pairs. If a trait's resemblance between twins is driven partly by genes, MZ pairs should be more similar than DZ pairs — and the size of that gap becomes the raw material for a heritability estimate. Twin researchers formalize this using the ACE model, which decomposes phenotypic variance (V_P) into three components:
where A is additive genetic variance, C is shared (common) environmental variance — environmental effects that make co-twins more alike (same house, same parenting, shared prenatal environment) — and E is non-shared (unique) environmental variance, including measurement error, that makes co-twins less alike despite shared genes and shared upbringing. Heritability, in the narrow sense, is h² = V_A / V_P, exactly as defined in Heritability: Broad and Narrow Sense; the twin design is simply a machine for estimating the sizes of A, C, and E from patterns of twin resemblance.
Mechanism — from twin biology to a heritability number
Step 1: Establish the two twin classes and their expected genetic sharing. MZ twinning occurs in roughly 3–4 per 1,000 births worldwide, at a rate that is largely constant across populations and appears developmentally stochastic rather than strongly heritable itself. DZ twinning rates vary considerably by population, maternal age, and fertility treatment (historically ~6–12 per 1,000 births, higher with assisted reproduction), because it depends on double ovulation, which does have a genetic and hormonal component. Crucially for the design:
| MZ twins | DZ twins | |
|---|---|---|
| Origin | one zygote splits | two zygotes, two eggs, two sperm |
| Expected allele sharing (IBD) | ~100% | ~50% (average) |
| Sex | always same-sex | same-sex or opposite-sex |
| Chorionicity | often monochorionic (shared placenta) | always dichorionic |
Step 2: Recruit twin pairs and measure the trait. Large twin registries (e.g., the Swedish Twin Registry, the Virginia Twin Registry, the UK's TwinsUK) recruit thousands of MZ and DZ pairs and measure a phenotype — a continuous trait like height, blood pressure, or IQ score, or a discrete disease status like schizophrenia or type 1 diabetes.
Step 3: Quantify twin resemblance.
- For continuous traits, compute the intraclass correlation coefficient (ICC) separately within MZ pairs (r_MZ) and DZ pairs (r_DZ).
- For discrete (disease) traits, compute concordance: the probability that the second twin has the disease given the first twin does. Two versions exist — pairwise concordance (concordant pairs / all pairs with at least one affected twin) and probandwise concordance (a proband-weighted version correcting for ascertainment bias, generally preferred).
Step 4: Apply the biometric logic. Under an additive genetic model with no dominance or epistasis, no gene-environment correlation, and the EEA:
(using A and C here as proportions of variance, with total variance normalized to 1, and E making up the remainder for MZ pairs since even genetically identical twins are not perfectly correlated). Subtracting:
This is Falconer's formula: h² ≈ 2(r_MZ − r_DZ). Once A is estimated, C = r_MZ − A, and E = 1 − r_MZ (since MZ twins share 100% of A and C, any remaining dissimilarity between them must be non-shared environment plus measurement error).
Step 5: Interpret the pattern of correlations.
- If r_MZ ≈ r_DZ (both similar and both substantial): suggests shared environment (C) dominates, with genes contributing little.
- If r_MZ ≈ 2 × r_DZ (DZ correlation about half the MZ correlation): consistent with a purely additive genetic model, C ≈ 0.
- If r_MZ > 2 × r_DZ (DZ correlation much less than half): suggests non-additive genetic effects (dominance, epistasis) are inflating MZ resemblance beyond the additive prediction, since MZ share dominance variance completely while DZ share only a quarter of it.
- If r_MZ is high but well below 1, the gap (1 − r_MZ) estimates E, the non-shared environmental (plus stochastic developmental and measurement) contribution — often the largest single component for many psychological traits.
Step 6: Formal model-fitting. In modern practice, researchers rarely just apply Falconer's algebraic formula; they fit the full ACE (or ADE, substituting dominance D for C) model to the variance-covariance matrices of MZ and DZ pairs using structural equation modeling or maximum likelihood, allowing formal significance testing of whether dropping A, C, or E worsens model fit.
The evidence
Francis Galton (1875) — the first twin study logic. Galton, in "The History of Twins as a Criterion of the Relative Powers of Nature and Nurture," was the first to propose comparing twins who were "closely alike" at birth to twins who were "not alike" as a way to disentangle heredity from environment, correctly intuiting that some twin pairs were more biologically similar than others (though he lacked the cytological concept of zygosity — MZ vs. DZ twins were not distinguished by biological mechanism until decades later, once chromosome behavior in meiosis and fertilization was understood). Galton's work is now understood to be entangled with his advocacy of eugenics, a historical fact addressed further below.
Curt Stern and colleagues (1920s–1930s) — establishing zygosity diagnosis. A prerequisite for a rigorous twin study is knowing which twin pairs are MZ and which are DZ, since same-sex DZ twins can look similar and be mistaken for MZ pairs. Researchers developed the "similarity method": comparing twins on many genetically influenced but environmentally neutral markers (blood groups, fingerprint ridge counts, and later many polymorphic blood and serum markers) to statistically infer zygosity, later validated and eventually superseded by direct DNA fingerprinting/genotyping, which gives a near-certain zygosity call from a handful of polymorphic markers.
Twin registry studies establishing schizophrenia heritability (mid-20th century, refined through the 1990s–2000s). A canonical demonstration of the twin method's power comes from schizophrenia concordance studies. Classic and modern pooled estimates consistently show:
| Relationship | Pairwise concordance for schizophrenia (approx.) |
|---|---|
| MZ twins | ~40–50% |
| DZ twins | ~10–15% |
| General population lifetime risk | ~1% |
The pattern is exactly what the twin logic predicts if genes matter substantially: MZ concordance is far above DZ concordance (roughly 3–4×, not 1×), and both are far above population base rate. But critically, MZ concordance is well below 100% — proving that genotype alone does not determine schizophrenia; substantial variance is attributable to non-shared environment (E), developmental noise, or gene-environment interaction. This dual conclusion — substantial heritability and substantial non-genetic contribution, both demonstrated in the same dataset — is one of the clearest empirical lessons twin studies deliver, and it directly rebuts the common misconception that a high heritability estimate implies environment is irrelevant.
The Minnesota Study of Twins Reared Apart (Bouchard et al., beginning 1979). This study specifically recruited MZ and DZ twin pairs who had been separated early in life and raised in different households, removing the shared-environment confound almost entirely (C should be near zero for twins reared apart, so any remaining resemblance is attributed to A + residual selective placement effects). Reared-apart MZ twins still showed substantial correlations for traits such as IQ (correlations around 0.7–0.8 in some analyses) and personality measures, correlations often nearly as high as those found in MZ twins reared together — a striking result that reduced-form twin theory predicted only if genetic effects, not shared upbringing, were driving most of the resemblance. The study also drew methodological criticism (non-random separation, media-driven recruitment of unusually similar-sounding pairs, and the assumption that "reared apart" fully equalized environments), illustrating that even this elegant design is not assumption-free.
Modern validation: comparing twin-based heritability to molecular/genomic estimates. A crucial independent check on the twin method has come from genome-wide approaches like GREML/GCTA, which estimate heritability from genome-wide SNP data in unrelated individuals — a completely different methodology sharing none of the twin design's assumptions about shared environment. For height, twin-based h² estimates (~0.8–0.9) and SNP-based heritability estimates (~0.4–0.5, rising toward twin estimates when rare/structural variants and better imputation are included) show reasonable convergence in rank-ordering and biological plausibility, though a persistent gap (the "missing heritability" problem) remains an active research area — itself a topic that leads directly into Quantitative Trait Loci and QTL Mapping, where researchers try to physically locate the genomic regions responsible for the A variance twin studies detect indirectly.
Worked examples
Example 1 — Falconer's formula applied to a hypothetical trait. Suppose a twin registry measures resting heart rate variability in 500 MZ pairs and 400 DZ pairs, and finds:
Apply Falconer's formula:
Then estimate shared environment:
A negative C is biologically impossible as a variance component; in practice this is reported as C ≈ 0, and the negative value is treated as evidence that the additive model fits poorly — specifically, r_DZ is less than half r_MZ (0.28 vs. 0.33 expected under pure additivity), which is the classic signature of non-additive genetic variance (dominance or epistasis), not literally negative shared environment. A more appropriate model here would be ADE (additive + dominance + unique environment) rather than ACE. Finally:
So roughly 34% of the variance in this trait is attributable to non-shared environmental and stochastic factors — a substantial fraction even under a model with strong genetic signal.
Example 2 — probandwise concordance and a clinical interpretation. A registry studies type 1 diabetes (T1D) in 160 MZ pairs and 420 DZ pairs where at least one twin is affected (the "proband"). Data:
| Pair type | Pairs ascertained | Concordant pairs (both affected) |
|---|---|---|
| MZ | 160 | 66 |
| DZ | 420 | 46 |
Pairwise concordance:
Interpretation: MZ concordance (41%) is roughly 3.7× DZ concordance (11%), and both are far above the general population incidence of T1D (~0.3–0.5% lifetime risk in most populations) — strong evidence of a substantial genetic contribution (consistent with known major contributions from HLA class II haplotypes plus polygenic background). But the fact that MZ concordance is only 41%, not near 100%, is clinically important: it tells a genetic counselor that even an identical twin of a T1D patient has a majority chance of never developing the disease, meaning environmental triggers (viral infection, gut microbiome, dietary factors implicated in ongoing research) or stochastic immune events are doing a large share of the causal work in whether disease actually manifests in a genetically susceptible individual. This distinction — high relative risk versus low absolute concordance — is exactly the kind of quantitative reasoning twin data are suited to support, and exactly the kind that gets flattened into overclaims ("T1D is genetic") in casual reporting.
Common misconceptions & traps
| Misconception | Why it's wrong |
|---|---|
| "A trait with h² = 0.8 is 80% caused by genes in each person." | Heritability is a population-level ratio of variances, not a per-individual causal share; it says nothing about any one person's development. |
| "High heritability means environment doesn't matter / can't be changed." | PKU has high heritability but is fully preventable by a dietary (environmental) intervention once detected; heritability estimated in one environment says nothing about effects of environments not sampled. |
| "MZ and DZ twins automatically experience equally similar environments." | The equal environments assumption can fail — MZ twins are often treated more similarly (same clothes, more shared friends, more confused-identity experiences), inflating h² if unaccounted for. |
| "Twin heritability tells us why groups (e.g., populations, sexes) differ in trait averages." | Within-population variance decomposition cannot be extrapolated to between-group mean differences, which may arise from entirely different (often purely environmental) causes — the classic "seed packets in different soil" analogy. |
| "Twin studies measure a fixed, universal number for each trait." | Heritability estimates are specific to the population, time, and range of environments sampled; the same trait can show different h² in different societies or eras (e.g., heritability of educational attainment differs across cohorts with different access to schooling). |
| "Concordance and correlation are interchangeable statistics." | Concordance (a proportion, for binary traits) and correlation (for continuous traits) are related but computed and interpreted differently; conflating them causes arithmetic errors in applying Falconer's formula. |
The scientist behind it
Origins · 2 minFrancis Galton and how this was discovered
Francis Galton, a half-cousin of Charles Darwin, was a restless Victorian gentleman-scientist who worked on subjects ranging from meteorology (he coined the term "anticyclone") to fingerprint classification to statistics, where he introduced the concepts of regression toward the mean and correlation. In 1875 he published the first systematic proposal to use twins to separate the effects of "nature" from "nurture," recognizing — from questionnaire data he solicited from twin families across Britain — that some twin pairs remained similar throughout life despite different life experiences, while other twin pairs who started out alike grew apart, and vice versa for twins who started dissimilar. Galton lacked the cytological vocabulary of monozygotic/dizygotic twinning (the chromosomal basis of heredity was not established until decades later), but his qualitative sorting of twins by degree of physical resemblance anticipated the MZ/DZ distinction. Galton is also, unavoidably, the founder of the eugenics movement (he coined the term in 1883), and his interest in heredity of "talent" and "character" was explicitly bound up with a program of selective breeding of humans that was later invoked to justify coercive and racist policies in multiple countries during the twentieth century. Modern behavioral and quantitative genetics use the statistical tools Galton pioneered while explicitly repudiating the eugenic program that motivated much of his original research question — a historical entanglement any serious treatment of twin-study heritability has to acknowledge rather than launder out of the record.
Where this matters today
Psychiatric genetics and risk counseling. Twin-derived heritability estimates for conditions like schizophrenia (~60–80%), bipolar disorder (~60–85%), autism spectrum disorder (~70–90%), and major depression (~30–40%) inform genetic counseling about recurrence risk in families and motivate the search for underlying risk variants via genome-wide association studies, a pipeline that continues into Quantitative Trait Loci and QTL Mapping for the biological loci that jointly produce the A variance component.
Distinguishing shared vs. non-shared environment in public health. Twin designs have repeatedly shown that for many behavioral and psychological traits, non-shared environment (E) — not shared family environment (C) — is the dominant environmental component, a finding (associated with Robert Plomin's work) that reoriented developmental psychology away from assuming that parenting style explains most sibling differences, and toward investigating peer groups, individual experiences, and stochastic developmental variation.
Complex disease epidemiology. Concordance studies in twins continue to be used to estimate heritability for common diseases (type 2 diabetes, obesity, cardiovascular disease, many cancers), providing benchmark heritability figures that geneticists use to judge how much of a disease's variance current genomic tools (polygenic risk scores) have managed to capture, and how much remains to be explained.
Distinguishing biology from confounding in epidemiology more broadly. The "co-twin control" design — comparing MZ twins who are discordant for an exposure (smoking, occupation, medication use) — is used as a natural quasi-experiment to control for genetic and shared-environment confounding when studying the effect of an environmental exposure on disease, complementing (and sometimes correcting) findings from ordinary observational epidemiology.
Agriculture and animal breeding (by extension). Although controlled breeding designs are the primary tool in livestock and crop genetics (unlike in humans), the same additive-variance logic underlying twin studies — comparing resemblance among relatives of known relatedness — is the conceptual ancestor of breeding-value estimation methods used to select high-heritability traits (milk yield, growth rate) for selective breeding programs.
Ask the tutor: How would a twin study's heritability estimate change if the environment became much more uniform across the whole population?
Do twin studies work for traits with no genetic basis at all?
Yes, and this is actually a useful validity check on the method. If a trait is purely environmentally determined with no genetic influence — for instance, which language a person's parents speak to them in infancy — MZ and DZ twins should show similar correlations (r_MZ ≈ r_DZ), both driven by shared upbringing, yielding an estimated h² near zero via Falconer's formula. Twin studies of such traits generally do produce near-zero heritability estimates, which is reassuring: the method isn't mechanically generating high heritability estimates for everything, and it correctly detects when a trait's resemblance is environmentally, not genetically, driven.
Can twin studies tell us whether a specific gene causes a trait?
No. Twin studies estimate the total proportion of variance attributable to all genetic differences combined, treating the genome as a black box; they identify no specific locus, gene, or variant. Locating the actual DNA regions responsible for the additive genetic variance twin studies detect is the job of linkage analysis and, more powerfully, genome-wide association studies and Quantitative Trait Loci and QTL Mapping, which scan the genome directly for statistical associations between variants and trait values. Twin studies and QTL/GWAS approaches are complementary: twin studies establish that genetic variance exists and roughly how much, while mapping studies try to find where in the genome it lives.
Why can heritability estimates differ for the same trait across different studies or countries?
Heritability is a ratio of variance components (V_A / V_P) specific to a particular population living in a particular range of environments at a particular time — it is not a fixed biological constant like a rate constant in chemistry. If environmental variance changes (say, a country equalizes access to nutrition or education), V_E shrinks relative to V_A, and heritability rises even though nothing about the genes or their effects has changed. This is why heritability of educational attainment, for example, differs between cohorts with unequal versus equal school access, and why heritability numbers from one country's twin registry shouldn't be assumed to transfer automatically to a different population or era.
What is the difference between pairwise and probandwise concordance, and why does it matter?
Pairwise concordance is simply (number of concordant pairs) / (total number of pairs with at least one affected twin). Probandwise concordance instead counts each affected individual as a separate "proband" and asks, for each proband, what proportion of their co-twins are also affected; concordant pairs are effectively counted twice (once per proband) in the numerator and denominator. Probandwise concordance is generally preferred because it corrects for ascertainment bias in how affected twin pairs get recruited into a study (pairs found through a single ill twin versus pairs found because both twins independently sought treatment), and it is the estimate directly comparable to individual disease risk in relatives.
Does a high heritability estimate mean a trait can't be changed by intervention?
No — this is one of the most consequential misconceptions in the field. Phenylketonuria (PKU) is a single-gene, highly heritable metabolic disorder, yet its devastating effect on intelligence is entirely preventable by a low-phenylalanine diet started in infancy, a purely environmental intervention. Heritability describes the current proportion of variance due to genetic differences within the environments actually sampled; it says nothing about what would happen if a novel intervention were introduced. Height is highly heritable within well-nourished modern populations, yet average heights rose substantially across the twentieth century due to nutritional and public health improvements — environmental change moved the population mean even though within-population heritability stayed high.
Are the Minnesota "twins reared apart" studies still considered reliable?
The Minnesota Study of Twins Reared Apart produced influential and broadly replicated findings — for instance, substantial IQ and personality correlations even in twins raised in very different households — and its core conclusion (that shared family environment explains less variance in adult personality/IQ than commonly assumed) has held up reasonably well against later, larger twin-registry work. However, methodologists have raised legitimate concerns: recruitment partly relied on media publicity, which may have selectively attracted unusually similar-seeming pairs; "reared apart" did not always mean radically different environments (some separated twins were placed with relatives in similar socioeconomic circumstances); and sample sizes for the rarest categories (DZ twins reared apart) were small. The study's qualitative conclusion is generally accepted, but its precise numerical estimates are treated with appropriate caution rather than taken as final values.
How does the twin method relate to newer genomic heritability estimates, and why don't they always match?
Genome-wide SNP-based methods (like GREML/GCTA) estimate heritability using measured genetic similarity among unrelated people, sidestepping the equal-environments assumption entirely. For many traits, SNP-based estimates come in noticeably lower than twin-based estimates — a gap dubbed "missing heritability." Leading explanations include: SNP arrays capture common variants well but miss rare variants, structural variants, and some non-additive genetic effects that twins fully share but unrelated individuals do not; and twin estimates may be somewhat inflated by unmodeled shared-environment or gene-environment correlation effects. The gap has narrowed as sequencing and modeling improve, and convergence between the two independent methods is treated as an important cross-validation of both.