Probability Basics and the Frequentist View
Flip a coin ten times and you might get seven heads. Does that mean the coin is biased? Ask a pregnant cow's veterinarian the chance of twins. She will quote a number built from thousands of past pregnancies, not from a single guess. Biology is a science of variable outcomes: which offspring survive, which allele a gamete carries, which patient responds to a drug. Probability is the mathematical language that lets biologists describe, predict, and test claims about that variability rigorously. This is rather than relying on intuition or anecdote. This article builds probability from the ground up using the frequentist interpretation: probability as the long-run relative frequency of an outcome across many repeated trials. This is the interpretation that underlies almost all classical Hypothesis Testing and Statistical Inference used in biology. It applies from Mendelian genetics to clinical trials. Understanding it precisely is essential. This means knowing what a "trial" is, what an "event" is. It also means grasping what it means for a probability to be a limit rather than a guarantee. This is the foundation on which every later rule is built. That includes addition, multiplication, conditional probability, and distributions. The big picture is simple but easy to misstate. Probability is not a claim about any single outcome. It is also not a prediction of what will happen in a small sample. A probability of 0.75 for a pea plant showing the dominant phenotype has a specific meaning. It does not mean "three out of the next four peas will be dominant." It means that if you grew an enormous number of peas from this cross, something specific would happen. The proportion showing the dominant phenotype would settle down, or converge, toward 0.75. Confusing the long-run limit with a short-run guarantee is a common error. It is the single most common conceptual error students make in biostatistics. It recurs throughout genetics, ecology, and epidemiology. This article delivers precise definitions of sample spaces and events. It offers a careful statement of the frequentist definition of probability with its axioms. It includes a worked demonstration using simulated genetic crosses. It covers the historical experiments that established this view of probability. Finally, it provides a guide to the misconceptions that trip up beginners. These are the errors beginners make before they reach the Addition Rule and Multiplication Rule.
- A sample space (S) lists every possible outcome of a random trial; an event is any subset of that space.
- The frequentist definition: P(A) is the limiting proportion of trials in which event A occurs, as the number of trials → ∞.
- Probabilities obey three axioms: non-negativity, total probability = 1, and additivity for disjoint events.
- A probability describes a population of possible trials, not a guarantee about any one trial or small sample.
- Real biological data (Mendel's pea counts, Buffon's coin flips) show sample proportions converging toward a stable theoretical value as sample size grows.
- This convergence is why probability underlies statistical testing: it justifies treating relative frequencies as estimates of true, fixed probabilities.
Probability is the value that a proportion of outcomes converges toward as the number of independent repeated trials grows without bound — it describes populations of trials, not single events.
The core idea
A random trial (or random experiment) is any process whose outcome cannot be predicted with certainty in advance but whose set of possible outcomes is known: flipping a coin, genotyping an offspring, measuring whether a patient recovers, sampling one individual from a population. The sample space, denoted (S), is the complete list of all possible outcomes of a single trial. For a coin flip, (S = \{H, T\}). For a cross between two heterozygous pea plants (Aa × Aa) at one locus, the sample space of genotypes is (S = \{AA, Aa, aA, aa\}), often collapsed by phenotype to (S = \{\text{dominant}, \text{recessive}\}). An event is any subset of the sample space — a collection of outcomes we choose to group together because we care about them. "The offspring shows the dominant phenotype" is the event ({AA, Aa, aA}). Events can be simple (a single outcome) or compound (several outcomes combined), and this distinction becomes essential once we combine events using AND/OR logic in the addition and multiplication rules that follow this topic. The frequentist definition of probability assigns a number P(A) to an event A by imagining the trial repeated an indefinitely large number of times, under identical conditions, independently each time. If event A occurs (n_A) times out of (n) total trials, the relative frequency is (n_A/n). The frequentist definition states:
Three axioms (due to Andrei Kolmogorov's 1933 formalization, though the intuition predates him by centuries) make this rigorous:
| Axiom | Statement |
|---|---|
| 1. Non-negativity | (P(A) \geq 0) for every event A |
| 2. Normalization | (P(S) = 1) — some outcome in the sample space always occurs |
| 3. Additivity | If A and B cannot both occur (mutually exclusive), (P(A \cup B) = P(A) + P(B)) |
Everything else in probability theory — the addition rule for overlapping events, the multiplication rule for independent events, conditional probability, and all downstream distributions used in biology (binomial, Poisson, normal) — is built from these three simple statements.
Mechanism — how frequentist probability actually works
Step 1: Define the trial precisely. Biological probability claims only make sense once the underlying repeatable trial is specified exactly. "Probability of twins in dairy cattle" implicitly means: among a well-defined population of pregnancies (a breed, a herd, a time period), what proportion result in twins? Change the population (a different breed, induced ovulation protocols) and the probability changes, because the sample space of "possible pregnancies" has changed.
Step 2: Enumerate the sample space. For discrete genetic crosses this is done with a Punnett square or, for multiple independent loci, by multiplying possibilities. A dihybrid cross (AaBb × AaBb) has a gamete sample space of 4 types per parent ((AB, Ab, aB, ab)), giving 16 equally likely zygote combinations.
Step 3: Assign probabilities to outcomes. Under Mendelian assumptions — independent segregation of alleles, equal viability of gametes — each of the 16 dihybrid combinations is equally likely, so each has probability 1/16. This is a theoretical probability, derived from a mechanistic model of meiosis (independent assortment of homologous chromosomes during Meiosis I), not yet an empirical measurement.
Step 4: Realize the trial repeatedly. Each actual mating, each actual offspring produced, is one realization of the trial. The relative frequency of a phenotype among real offspring is a sample statistic — an estimate, not the true probability itself.
Step 5: Watch convergence. As the number of offspring scored grows — 10, then 100, then 10,000 — the observed relative frequency wobbles less and settles near the theoretical value (e.g., 9/16 for the dominant-dominant phenotype class). This settling-down is the empirical signature of a real, fixed underlying probability, and it is what the frequentist definition formally describes as a limit.
Step 6: Distinguish the limit from any one sample. With only 16 offspring, getting 12 dominant-dominant instead of the expected 9 is unsurprising — sampling variability is large at small (n). Biologists use this understanding directly when they build Hypothesis Testing procedures like the chi-square goodness-of-fit test: a specific deviation from the expected ratio is judged "surprising" or "not surprising" only in light of how much variability is expected at that sample size.
The evidence — experiments that established the frequentist view
Buffon's coin-tossing and needle experiments (Georges-Louis Leclerc, Comte de Buffon, mid-1700s). Buffon personally tossed a coin 4,040 times and recorded 2,048 heads, a relative frequency of 0.5069 — very close to the theoretical 0.5 expected under a fair coin, illustrating stabilization of relative frequency with a large number of trials. This is one of the earliest documented large-scale empirical demonstrations that repeated trials of a chance process converge toward a stable proportion, the core empirical claim of frequentism. Later, the statistician Karl Pearson had an assistant toss a coin 24,000 times, recording 12,012 heads (relative frequency 0.5005), further reinforcing the pattern with even larger (n).
Mendel's pea-plant crosses (Gregor Mendel, 1856–1863). Mendel's experiments are the founding evidence base for treating heredity probabilistically. In his monohybrid crosses of pea plants heterozygous for seed shape (Rr × Rr), Mendel scored 7,324 offspring seeds and observed 5,474 round and 1,850 wrinkled, a ratio of 2.96:1 — remarkably close to the theoretical 3:1 ratio predicted by independent segregation of alleles at meiosis. Crucially, Mendel did not get exactly 3:1 in any single small batch; individual plants' seed counts deviated substantially from 3:1, but the aggregate across thousands of seeds converged tightly. This is the frequentist definition in action: the true probability (3/4 round, given the underlying genetic mechanism) is only revealed reliably by accumulating a very large number of trials, exactly as the limit definition requires.
R.A. Fisher's 1936 statistical reanalysis of Mendel's data famously showed Mendel's results were, across all his experiments, suspiciously close to the expected ratios — closer than typical sampling variability would predict by chance (a chi-square analysis suggesting the fit was "too good"). This controversy (still debated) does not undermine the frequentist logic; if anything, it depends on it: Fisher's argument works only because we know what range of relative frequencies should occur around the true probability due to ordinary sampling variation, and Mendel's aggregate results fell outside that expected range on the "too good" side.
What these experiments jointly proved: relative frequencies of repeatable biological and physical trials are not arbitrary — they stabilize near a fixed value as sample size grows, that value can be predicted from an underlying causal mechanism (fair coin geometry; equal segregation of alleles at meiosis), and deviations from the expected proportion shrink, on average, as (1/\sqrt{n}) with increasing trials. This last quantitative fact is what later justifies the standard error formulas used throughout biostatistics.
Worked examples
Example 1 — a dihybrid cross with asymmetric linkage disruption. Suppose two loci in a fictional model organism are normally unlinked (independent assortment), giving the classic dihybrid 9:3:3:1 phenotype ratio in an (AaBb \times AaBb) cross. A researcher crosses AaBb × AaBb and examines 1,600 offspring. Under the null model of independent assortment:
Expected counts among 1,600 offspring: 900, 300, 300, 100. Observed counts: 860, 305, 340, 95. Relative frequencies observed: 860/1600 = 0.5375, 305/1600 = 0.1906, 340/1600 = 0.2125, 95/1600 = 0.0594. The (aaB\_) class is noticeably inflated (0.2125 vs. expected 0.1875) and the double-dominant class is depressed (0.5375 vs. 0.5625). This pattern — a shortage of one parental-type combination and excess of a recombinant combination — is the classic frequentist signature that the sample space assumption ("all 16 gamete combinations equally likely, loci independent") may be wrong; the loci may be genetically linked rather than independently assorting. Note what the frequentist framework lets us do here: we compare an observed relative frequency to a theoretical one and ask whether the discrepancy is larger than sampling variability alone would produce — the beginning of a chi-square goodness-of-fit test, formalized once you reach The Chi-Square Test.
Example 2 — clinical relative frequency and sample size, a diagnostic test scenario. A new rapid diagnostic test for a bacterial infection is trialed on patients with confirmed infection (established by gold-standard culture). In a pilot study of 20 confirmed-positive patients, the test correctly flags 17 as positive, an observed sensitivity (true positive rate) of 17/20 = 0.850. In a larger multi-center validation of 2,400 confirmed-positive patients, the test correctly flags 1,932, an observed sensitivity of 1932/2400 = 0.805. A clinician asks: which number, 0.850 or 0.805, is the "true" sensitivity of the test? The frequentist interpretation gives a clear answer: neither observed proportion is the true sensitivity — each is an estimate of a fixed underlying probability, obtained from a finite number of trials. The larger study (n = 2,400) gives a more reliable estimate because relative frequency estimates from larger samples have smaller sampling variability (their standard error shrinks roughly as (1/\sqrt{n})); the pilot's estimate of 0.850, based on only 20 patients, is subject to much larger random fluctuation and should not be trusted over the larger study's 0.805. This is precisely why regulatory bodies (e.g., the FDA) require large validation cohorts before approving diagnostic tests: they are explicitly relying on the frequentist convergence property to pin down the true sensitivity within an acceptably narrow margin.
Common misconceptions & traps
| Misconception | Why it's wrong |
|---|---|
| "P = 0.75 means 3 of the next 4 offspring will show the trait" | Probability describes the long-run limit across huge numbers of trials, not a guarantee for any small, finite sample. |
| "A rare event just happened, so it's 'due' to balance out soon" (gambler's fallacy) | Independent trials have no memory; a run of recessive offspring does not make dominant offspring more likely next time. |
| "If P(A) = 0.5, then A and 'not A' are automatically equally likely to have caused any given outcome" | The probability applies to the event definition, not to any inference about mechanism after the fact; this conflates probability with causal explanation. |
| "The sample space must always list equally likely outcomes" | Equal likelihood is a special assumption (true for fair alleles under Mendelian segregation) — not a universal property of sample spaces; outcomes like "survive" vs. "die" are rarely equally likely. |
| "Observing a proportion close to a hypothesized value in one sample proves the hypothesis is correct" | Any single sample's relative frequency is subject to sampling error; only accumulated evidence and formal hypothesis testing can support or reject the hypothesis. |
| "Probability is a property of a single individual trial (e.g., this specific patient has a 30% chance of relapse, full stop)" | Frequentist probability strictly refers to a proportion across a defined reference class of repeated trials, not an intrinsic property of one unrepeated case — a subtlety important in clinical risk communication. |
The scientist behind it
The frequentist view of probability, though anticipated by earlier gambling mathematicians (Jacob Bernoulli's Ars Conjectandi, 1713, proved the Law of Large Numbers — the mathematical theorem underlying convergence of relative frequency), was given its rigorous modern axiomatic foundation by the Russian mathematician Andrei Kolmogorov in 1933, and its application to biological data was pioneered decisively by Gregor Mendel, whose meticulous, large-scale counting of pea-plant offspring turned heredity from a qualitative art into a quantitative, probabilistic science. Mendel was an Augustinian friar and trained physicist/botanist working in the monastery garden at Brno (in the modern Czech Republic). Between 1856 and 1863 he grew and hand-pollinated tens of thousands of pea plants (Pisum sativum), tracking seven clearly binary traits (seed shape, seed color, pod shape, flower position, and others) across multiple generations. His crucial methodological innovation was scale and precision: rather than describing hybrid offspring qualitatively ("mostly like one parent"), he counted every single seed and plant, generation after generation, and computed exact ratios. This let him notice that ratios like 3:1 and 9:3:3:1 recurred reliably — a discovery invisible without large-(n) counting, and impossible to trust without an implicit frequentist logic that large samples reveal true underlying proportions.
Origins · 1 minGregor Johann Mendel and how this was discovered
Where this matters today
Clinical trial design. Regulatory approval of drugs and vaccines depends on estimating true response rates and adverse-event rates from large trial cohorts, explicitly relying on the frequentist principle that bigger samples yield relative frequencies closer to the true population probability — this is why phase III trials enroll thousands of patients rather than dozens.
Genetic counseling and prenatal screening. Quoting a patient a "1 in 4 chance" of a recessive genetic disorder for a given pregnancy is a direct, real-time application of Mendelian probability calculated from sample-space reasoning at the gamete level, tracing straight back to Mendel's ratios.
Conservation biology and population viability analysis. Estimating extinction probabilities for endangered species uses long-run frequency reasoning applied to simulated or historical population trajectories — a single population's fate is uncertain, but probabilities are estimated across many simulated or comparable populations.
Epidemiological surveillance. Disease prevalence and incidence rates reported by public health agencies (e.g., CDC, WHO) are relative frequencies computed from large surveyed or tested populations, interpreted as estimates of an underlying true probability of infection or disease in the population.
Diagnostic test validation and biotechnology quality control. Manufacturing defect rates, false-positive/false-negative rates for lab assays, and genome-sequencing error rates are all estimated as long-run relative frequencies from large numbers of repeated measurements, forming the statistical backbone of quality assurance in biotech pipelines.
Agricultural breeding programs. Predicted trait ratios in crop and livestock breeding crosses (e.g., disease-resistance alleles) rely on the same sample-space and probability machinery introduced by Mendel, scaled up with modern marker-assisted selection.
Ask the tutor: If a genetic cross gives a ratio close to but not exactly 3:1 in a moderate-sized sample, how do we decide whether the deviation is just sampling variability or evidence of a real biological effect (like linkage or reduced viability)?
What exactly counts as a "trial" in biology, if organisms and experiments are never perfectly identical?
In practice, biologists define a trial as a class of events considered interchangeable for the purposes of the question at hand — e.g., "a gamete produced by an Aa heterozygote," or "a randomly selected patient with this diagnosis in this hospital system." No two organisms are literally identical, but the frequentist framework only requires that the relevant probability-generating mechanism (allele segregation, disease biology) behaves consistently across the trials being pooled. This is why defining the reference population precisely — the species, the genotype, the clinical subgroup — matters enormously; changing the reference class changes the probability, which is why genetic counseling always specifies exact parental genotypes before quoting a risk.
Is the frequentist definition the only way to define probability in biology?
No. The Bayesian interpretation treats probability as a degree of belief that can be updated with new evidence, even for one-off, non-repeatable events (e.g., "the probability this newly discovered fossil belongs to species X"), and is increasingly used in phylogenetics and genomics. The frequentist view, however, remains the default foundation for classical hypothesis testing, p-values, and confidence intervals taught first in biostatistics, because it maps directly onto the idea of repeated experiments or repeated sampling from a population — the situation most common in controlled biological experiments and large surveys.
Why can't we just calculate probability by counting outcomes in the sample space, as with the Punnett square, without doing any actual experiment?
That calculation gives a theoretical probability, derived from an assumed model (e.g., equal segregation of alleles, no linkage, no differential mortality). It is extremely useful for prediction, but it is only as good as the model's assumptions. The frequentist, empirical approach — actually running the cross and counting real offspring — is essential because real biological systems can violate the assumptions (genetic linkage, segregation distortion, differential embryo viability). Comparing observed relative frequencies to theoretical ones, as in the dihybrid-cross worked example, is exactly how such violations are detected.
Does a probability of zero mean an event is impossible, and probability of one mean it's certain?
For discrete sample spaces with a finite number of outcomes (like genotype classes), yes — P(A) = 0 means A cannot occur given the model, and P(A) = 1 means A must occur. For continuous sample spaces (e.g., exact body mass measured with infinite precision), a single exact value technically has probability zero even though it isn't "impossible" in the everyday sense; this subtlety is addressed later with continuous probability distributions and density functions, and does not affect the discrete genetics and epidemiology examples used at this level.
How does the Law of Large Numbers relate to the frequentist definition?
The Law of Large Numbers, proven by Jacob Bernoulli in 1713, is the mathematical theorem that guarantees relative frequency converges to the true probability as the number of independent, identically distributed trials grows without bound. The frequentist definition is essentially the informal, applied statement of this theorem: it tells biologists that if they run enough replicate trials — enough offspring, enough patients, enough independent samples — their observed proportion will reliably approach the true underlying probability, which is precisely why Mendel's large-scale seed counts, and modern large clinical trials, are trustworthy in a way small pilot studies are not.
How does this connect to what comes next — the addition and multiplication rules?
Once sample spaces, events, and the frequentist meaning of P(A) are established, the natural next questions are how to combine events: what is the probability that A or B occurs (the Addition Rule and Mutually Exclusive Events), and what is the probability that A and B both occur (the Multiplication Rule and Independence)? Both rules are built directly from the three axioms introduced here — additivity for disjoint events underlies the addition rule, while the definition of independent, repeated trials underlies the multiplication rule used for combining independent genetic loci or independent diagnostic tests.
Can probability ever be estimated from just one organism or one measurement?
No single measurement can estimate a probability on its own — a probability describes a proportion across many potential trials, so at least a modest sample is required to get any meaningful estimate, and larger samples give more precise, more trustworthy estimates. This is precisely the logic behind minimum sample size requirements in experimental biology: a genetics cross with only 4 offspring cannot meaningfully test a 3:1 ratio hypothesis, while one with 400 offspring can, because the relative frequency in the larger sample is far less subject to random sampling swings.