Independence Revisited
Two loci sit on the same chromosome, ten map units apart. Do they get inherited independently? A student's intuition says every gene assorts independently. That is what an early biology class taught. The intuition is wrong. The reason it is wrong is worth making precise. Independence is not a biological default. It is a formal statistical condition. Some pairs of events satisfy it, and many do not. This article gives that condition a rigorous face: P(A|B) = P(A), equivalently P(A ∩ B) = P(A)·P(B). It then asks the more interesting question. When does real biology break this equation, and how do we detect the break? Linked genes violate independence because physical proximity on a chromosome makes recombination between them rare. Knowing one allele changes the odds on the other. Infectious disease violates independence because one person's infection status changes their neighbor's risk. That is exactly why epidemic models cannot just multiply individual probabilities together. The payoff is conceptual leverage that reaches into nearly every quantitative model in biology. The Binomial Distribution assumes independent trials; The Poisson Distribution assumes independent rare events in non-overlapping intervals. Every time those models are applied — to mutations, to births, to random mating — independence is the hidden assumption doing the work. This article teaches you to test that assumption instead of taking it on faith. It uses chi-square-style reasoning, linkage data, and the mathematics of contagion.
- Independence means P(A|B) = P(A): learning B happened tells you nothing new about A.
- Equivalently, independent events satisfy P(A ∩ B) = P(A) × P(B) — a multiplication shortcut that only works under independence.
- Genes on different chromosomes (or far apart on the same one) assort independently; genes close together on the same chromosome are linked and violate independence.
- Recombination frequency (map distance) is a direct, measurable departure from the 50% independence prediction.
- Disease transmission is the classic biological case where events are not independent — one infection changes the probability of the next.
- A chi-square test for independence is the standard tool for deciding, from real data, whether two biological variables truly behave independently.
Independence is a testable numerical claim — P(A|B) = P(A) — not a biological assumption to be taken for granted, and linkage, contagion, and shared ancestry are exactly the situations where it fails.
The core idea
Two events A and B are statistically independent if knowing that B occurred does not change the probability of A. Formally,
P(A | B) = P(A) whenever P(B) > 0. Because conditional probability is defined as P(A|B) = P(A ∩ B)/P(B), this condition rearranges into the more commonly quoted product rule for independent events:
P(A ∩ B) = P(A) × P(B) These two statements are logically equivalent, and either can serve as the operational definition. The product-rule form is usually easier to check against data because it involves only unconditional probabilities that can be estimated directly by counting. It helps to separate three ideas that students often blur together. Independence is a probabilistic property of two events, defined by the equation above. Mutual exclusivity (disjointness) is a different, often confused property: A and B are mutually exclusive if they cannot both happen, meaning P(A ∩ B) = 0. Mutually exclusive events with nonzero individual probabilities are in fact maximally dependent — knowing B occurred tells you A definitely did not, which is about as far from P(A|B) = P(A) as two events can get. Correlation is yet another idea, a linear-association measure for numeric variables; two variables can be dependent (in the probability sense) while having zero linear correlation, if their relationship is nonlinear. In biology it is dependence, not correlation, that maps onto the deep mechanistic question of whether one event's outcome physically or informationally constrains another's. The practical value of the independence test is that it is falsifiable with real counts. Given contingency data — allele at locus 1 crossed with allele at locus 2, or infection status of person 1 crossed with infection status of their household contact — you compute the observed joint frequency and compare it to the product of the marginal frequencies predicted under independence. Large, systematic discrepancies are the statistical fingerprint of an underlying mechanism: physical linkage on a chromosome, or person-to-person transmission of a pathogen.
Mechanism — how the test actually works, and how biology breaks it
Step 1: State the independence prediction. For two genetic loci assorting independently, Mendel's second law predicts that the four gamete classes from a dihybrid (AaBb) — AB, Ab, aB, ab — occur with probability 1/4 each, because P(A allele) and P(B allele) multiply freely. Any deviation from that 1:1:1:1 gamete ratio (or the corresponding 9:3:3:1 F₂ phenotype ratio) is evidence against independent assortment.
Step 2: Understand the physical mechanism that produces independence — or breaks it. Independent assortment holds when two loci sit on different chromosomes, because homologous chromosome pairs orient randomly at metaphase I of meiosis; which chromosome from a pair goes to which pole is uncorrelated with the orientation of any other pair. This is a direct mechanical consequence of spindle attachment being independent across bivalents. When two loci sit on the same chromosome, this mechanical randomization no longer applies automatically. The alleles travel together into the same gamete unless a crossover (physical exchange of chromosome segments during prophase I) occurs between them. The closer two loci are physically, the less chromosome length lies between them, and the lower the chance a crossover falls in that interval. Recombination frequency — the proportion of recombinant (non-parental) gametes — is therefore a direct, measurable readout of physical distance, and it approaches (but rarely exceeds) 50% only for loci far enough apart, or on different chromosomes, to behave as if independent.
Step 3: Quantify the departure. Define recombination frequency:
r = (number of recombinant gametes) / (total gametes)
- r = 0.50 → the two loci behave as statistically independent (unlinked, or so far apart that crossovers randomize them).
- r < 0.50 → the loci are linked; parental combinations are systematically overrepresented, which is precisely a violation of P(A ∩ B) = P(A)·P(B).
- r near 0 → the loci are essentially always inherited together (tight linkage).
One map unit (centimorgan, cM) is defined as 1% recombination frequency, giving a direct ruler for genetic distance along a chromosome.
Step 4: The same logic, applied to disease. In an independence world, if the population prevalence of infection is p, then the probability that two random individuals are both infected would simply be p². But infectious disease transmission is mechanistically a contact process: pathogen particles move from an infected host to a susceptible host during physical or respiratory contact, converting susceptible individuals into infected ones at a rate proportional to the number of infectious contacts. Because household members, sexual partners, and classroom neighbors share contact networks, one person's infection status directly raises the conditional probability that their contacts are infected. That is exactly P(infected₂ | infected₁) > P(infected₂) — a violation of independence built into the transmission mechanism itself, and it is why epidemic models use compartmental (SIR-type) frameworks rather than simple binomial products.
The evidence — how linkage was proved to violate independence
The classic demonstration comes from Thomas Hunt Morgan's laboratory work on Drosophila melanogaster, beginning around 1910–1913, extended quantitatively by his undergraduate student Alfred Sturtevant.
The observation. Morgan crossed fruit flies heterozygous for two X-linked traits — body color (gray/black) and wing shape (normal/vestigial) — and examined the offspring's phenotype ratios. Under Mendel's law of independent assortment, a dihybrid testcross should yield the four phenotypic classes in a 1:1:1:1 ratio. Morgan instead found the two parental combinations (gray-normal and black-vestigial) vastly overrepresented, and the two recombinant combinations (gray-vestigial and black-normal) rare — a pattern utterly inconsistent with independence.
What it proved. The skewed ratio showed that the two genes were not assorting independently: knowing the body-color allele changed the probability of the wing-shape allele, exactly the P(A|B) ≠ P(A) signature. Morgan correctly inferred that genes physically located on the same chromosome are inherited together unless separated by crossing-over, and that the frequency of the rare recombinant class is a quantitative measure of how far apart the two genes lie.
Sturtevant's extension (1913). As an undergraduate, Sturtevant realized that if crossover frequency scales with physical distance, one could use recombination frequencies between many gene pairs to build a linear map — the first genetic linkage map ever constructed, for six X-linked Drosophila genes. Genes with high recombination frequency between them were placed far apart; genes with very low recombination frequency were placed close together; and the map units (later named centimorgans in Morgan's honor) were additive along the chromosome, just as physical distances are additive along a line. The construction only works because linked loci are not independent, and the degree of non-independence is a smooth, measurable function of physical distance — turning a statistical anomaly into a quantitative mapping tool that remains in use in modified forms (LOD scores, recombination-based genetic maps) in modern genomics.
A complementary line of evidence came later from cytological observation: Curt Stern (1931) and, independently, Harriet Creighton and Barbara McClintock (1931, in maize) showed that recombination of genetic markers is physically accompanied by exchange of visibly distinguishable chromosome segments under the microscope — directly tying the statistical departure from independence to an observable physical event (crossing-over), not merely inferring it from ratios.
Worked examples
Example 1 — A linkage cross with asymmetric numbers. In maize, consider two linked loci: C (colored aleurone, dominant) and Sh (full/shrunken seed, dominant), both on chromosome 9. A heterozygous plant CcShsh (in coupling: dominant alleles inherited together, C Sh / c sh) is testcrossed to ccshsh. Suppose 8,000 offspring are scored:
| Phenotype class | Observed | Type |
|---|---|---|
| Colored, full (C Sh) | 3,580 | parental |
| Colorless, shrunken (c sh) | 3,620 | parental |
| Colored, shrunken (C sh) | 420 | recombinant |
| Colorless, full (c Sh) | 380 | recombinant |
Recombination frequency: r = (420 + 380) / 8000 = 800/8000 = 0.10 = 10 cM Under independence, we would expect all four classes at 2,000 each (25%), giving P(C ∩ Sh) = P(C)·P(Sh) = 0.5 × 0.5 = 0.25. Instead, P(C ∩ Sh) ≈ 3580/8000 = 0.4475, wildly above the independence prediction of 0.25. The two loci are 10 map units apart — moderately linked, close enough that parental combinations dominate but far enough that a real, measurable 10% recombinant class still appears. A chi-square goodness-of-fit test against the 1:1:1:1 independence expectation (df = 3) would return a chi-square value in the thousands here, decisively rejecting independence (critical value at α = 0.05 is only 7.815).
Example 2 — Contagion and the failure of the product rule in an outbreak. A university dormitory has 200 students. Suppose the semester-long probability that any given, isolated student contracts a particular norovirus strain, absent any local transmission, is estimated at p = 0.05 from historical baseline surveillance (community background rate). If infections were independent across roommates, the probability that a specific pair of roommates are both infected would simply be: P(both infected | independence) = 0.05 × 0.05 = 0.0025 (0.25%) But actual surveillance data from the outbreak show that among 60 roommate pairs where the first roommate became infected, 27 of the second roommates also became infected — a conditional attack rate of: P(second infected | first infected) = 27/60 = 0.45 (45%) This is 180-fold higher than the 0.05 baseline rate, and vastly higher than the 0.0025 joint probability predicted under independence. The mechanism is direct: shared bathrooms, shared air, and close contact transmit norovirus particles from an infected roommate to a susceptible one, physically coupling the two individuals' infection probabilities. This single number — P(A|B) ≫ P(A) — is the quantitative signature epidemiologists use to detect an outbreak (local clustering) rather than a coincidental rise in independent background cases, and it is exactly why the secondary attack rate (conditional probability of transmission to a contact) is a core surveillance statistic, distinct from the population prevalence.
Common misconceptions & traps
| Misconception | Why it's wrong | |
|---|---|---|
| "Mutually exclusive events are a kind of independent events." | They are nearly opposite: if A and B are mutually exclusive with P(A), P(B) > 0, then P(A\ | B) = 0 ≠ P(A), so they are maximally dependent. |
| "All genes assort independently — that's Mendel's law." | Mendel's second law strictly applies to loci on different chromosomes (or very far apart on the same one); physically close loci are linked and violate independence. | |
| "If two variables show zero correlation, they must be independent." | Zero linear correlation only rules out a linear relationship; strongly nonlinear dependence can still exist with near-zero correlation coefficient. | |
| "A high joint probability P(A∩B) automatically implies dependence." | You must compare P(A∩B) to the product P(A)·P(B), not just look at its raw size — two common events can still be independent if the product matches. | |
| "Independence in probability is the same as causal independence." | Two events can be statistically associated (dependent) without either causing the other — e.g., both driven by a shared confounder such as ancestry or population structure. |
The scientist behind it
Origins · 1 minThomas Hunt Morgan and how this was discovered
Morgan began his career skeptical of Mendelian and chromosomal theories of heredity, initially favoring environmental explanations for variation. His conversion came from his own data: a spontaneous white-eyed male fruit fly appeared in his Drosophila cultures in 1910, and its inheritance pattern tracked the X chromosome with uncanny precision. Pursuing this observation with an assembly line of student researchers in Columbia's famously cramped "Fly Room" — including Alfred Sturtevant, Calvin Bridges, and Hermann Muller — Morgan generated the dihybrid cross data that revealed systematic, non-independent inheritance of X-linked traits. He correctly interpreted this as physical linkage along a chromosome, disrupted only by crossing-over, and his lab's quantitative recombination data became the empirical foundation for chromosomal mapping. Morgan received the Nobel Prize in Physiology or Medicine in 1933 "for his discoveries concerning the role played by the chromosome in heredity," the first Nobel awarded for work in genetics.
Where this matters today
Clinical genetics and linkage mapping. Modern human disease-gene mapping still exploits departures from independence: markers that co-segregate with a disease phenotype more often than chance (a low recombination frequency, quantified via LOD scores) point investigators toward the physical location of causal genes, a strategy that has localized loci for conditions from cystic fibrosis to Huntington disease before full genome sequencing was routine.
Pharmacogenomics and haplotype blocks. Because nearby single-nucleotide polymorphisms (SNPs) are inherited together in linkage disequilibrium — a population-genetics generalization of the same non-independence principle — genome-wide association studies can flag disease risk using a handful of tagged SNPs rather than sequencing every base, dramatically cutting genotyping cost.
Infectious disease surveillance and outbreak detection. Public health agencies routinely compute secondary attack rates and compare them to baseline prevalence precisely to detect the P(A|B) ≠ P(A) signature of transmission, distinguishing genuine outbreaks (contact-driven, non-independent case clustering) from a random uptick in unrelated cases. This logic underlies contact tracing protocols for tuberculosis, measles, and COVID-19.
Conservation genetics. Non-independence of alleles caused by small population size (inbreeding) — where relatedness makes the two alleles at a locus in one individual not statistically independent draws from the population gene pool — is quantified by the inbreeding coefficient F, informing captive breeding programs for endangered species to avoid deleterious homozygosity.
Agricultural breeding. Marker-assisted selection in crop and livestock breeding relies on knowing which trait loci are linked (violate independence) versus unlinked, so that breeders can predict which desirable trait combinations will remain together across generations and which will be broken apart by recombination.
Ask the tutor: If two loci show 50% recombination frequency, does that prove they are on different chromosomes — or could they still be linked?
What's the difference between independence and "no correlation"?
Independence is a strong condition: P(A|B) = P(A) must hold for every value of the variables, ruling out any relationship, linear or not. Zero correlation only means there is no linear association — two variables can be perfectly dependent through a nonlinear (e.g., U-shaped) relationship while still showing a correlation coefficient of zero. In biology, dose-response relationships are a classic example: toxin concentration and cell survival can have a strongly curved, highly dependent relationship that a linear correlation coefficient badly understates. So independence implies zero correlation, but zero correlation does not imply independence.
Why do genes on different chromosomes assort independently?
During meiosis I, each homologous chromosome pair (bivalent) orients randomly at the metaphase plate, and this orientation is mechanically uncoupled from how any other bivalent orients — spindle fiber attachment to one pair does not influence attachment to another. Because which chromosome goes to which pole is effectively a separate coin flip for each pair, alleles on different chromosomes end up in gametes according to independent random processes, reproducing the P(A∩B) = P(A)·P(B) rule exactly. This is the cellular, mechanical basis for Mendel's second law, and it breaks down only when two loci share a single chromosome and crossing-over is incomplete.
Can two genes on the SAME chromosome ever behave as independent?
Yes — if they are far enough apart, the probability of at least one crossover occurring somewhere between them during meiosis approaches certainty, and multiple crossovers can occur and cancel out, randomizing the allele combinations passed to gametes. In practice, recombination frequency asymptotically approaches 50%, the same value expected under true independence, once two loci are far apart enough (often >50 map units, sometimes on opposite arms of a long chromosome). This is why classical linkage mapping caps out at about 50% apparent recombination even for genes known to be on the same physical chromosome.
How do biologists formally test independence with real data?
The standard tool is a chi-square test of independence, applied to a contingency table of observed joint counts. You compute expected counts under the independence assumption (row total × column total ÷ grand total for each cell), then calculate χ² = Σ (observed − expected)² / expected. A large χ² relative to the chi-square distribution's critical value (given the appropriate degrees of freedom) leads to rejecting independence. In genetics this is applied directly to gamete or offspring class counts from a testcross; in epidemiology it is applied to exposure-by-outcome tables.
Does independence ever hold exactly in biology, or only approximately?
Almost always approximately, because biological systems are full of shared causes — shared ancestry, shared environment, shared chromosomes — that induce at least weak dependence somewhere. Unlinked genes on different chromosomes come very close to true independence because meiotic segregation is a highly reliable, near-mechanical randomizer. But truly perfect independence is an idealization, useful because it gives a clean null hypothesis against which real, imperfect biological data can be tested and its departures quantified (as recombination frequency, secondary attack rate, or a chi-square statistic).
How does this connect to the Binomial and Poisson distributions I'll see next?
Both distributions are built on an independence assumption at their core. The binomial distribution assumes a fixed number of trials that are independent and identically distributed (each with the same success probability), so that the probability of any particular sequence of successes and failures is a simple product; violating independence (e.g., contagious events, or genetically linked traits) breaks the binomial's variance formula and its predicted probabilities. The Poisson distribution similarly assumes that events occur independently in continuous time or space, which is why clustering processes (like outbreaks or spatially clumped mutations) are famously poorly fit by a plain Poisson model and instead require overdispersed alternatives.
Is linkage disequilibrium the same thing as physical linkage?
Not quite — they are related but distinct. Physical linkage refers to two loci being close together on the same chromosome, which reduces recombination frequency below 50%. Linkage disequilibrium (LD) is the population-genetics statistic measuring whether specific alleles at two loci co-occur more or less often than expected under independence, given their individual allele frequencies in the population. Physically linked loci often show high LD because recombination hasn't had enough generations to shuffle them apart, but LD can also arise from population history (recent admixture, selection, or a population bottleneck) even between unlinked loci, so LD is really the direct population-level analogue of the P(A∩B) ≠ P(A)·P(B) test applied to alleles.