Parameters versus Statistics

A population is the complete set of individuals, measurements, or units a biologist wants to draw conclusions about — every largemouth bass in a lake, every cell in a tumor, every human born with a particular mutation. A parameter is a numerical summary of some measurable characteristic of that entire population: the true mean body mass of all the bass, the true proportion of tumor cells expressing a marker, the true standard deviation of litter size across all foxes in a region. Parameters are constants. They have one fixed true value at any given time, even though we almost never know what that value is, because measuring an entire population is usually impossible, impractical, or destructive. A sample is a subset of the population that is actually measured. A statistic is a numerical summary calculated from that sample — the mean mass of the 40 bass you caught and weighed, the proportion of 200 tumor cells stained positive under the microscope, the standard deviation of litter size across the 25 fox dens you tracked. Unlike a parameter, a statistic is not fixed: it is a number you compute, and it depends entirely on which particular individuals ended up in your sample. Catch a different 40 bass and you will almost certainly get a slightly different mean mass, even if nothing about the lake's fish population changed at all. The entire logical structure of inferential statistics depends on keeping these two categories separate. A statistic is not merely a smaller version of a parameter — it is a random variable, because its value depends on the luck of which sample you happened to draw. A parameter, by contrast, has no randomness in it at all: it is simply unknown to us. The practice of biostatistics is built almost entirely on the project of using known, variable statistics to make honest, quantified guesses about unknown, fixed parameters. This is the conceptual seed for confidence intervals, hypothesis tests, and standard errors — all of which exist to describe how much a statistic is likely to differ from the parameter it is estimating.

Step 1: Define the population and the parameter of interest. Before any data are collected, a researcher must specify exactly what population they care about (e.g., "all adult zebrafish in this aquaculture facility") and what numerical feature of it they want to know (e.g., the true mean body length, denoted μ). This parameter exists as a single fixed number the moment the population exists, whether or not anyone ever measures it.

Step 2: Draw a sample. Because measuring every individual is rarely feasible, the researcher selects a subset — ideally by a method such as simple random sampling, in which every individual in the population has an equal and independent chance of being chosen. Sample size is denoted n; population size is denoted N. Typically n ≪ N.

Step 3: Compute the statistic from the sample. Using the same formula the true parameter would require, but applied only to the sampled individuals, the researcher computes the corresponding statistic — the sample mean x̄, the sample standard deviation s, the sample proportion p̂. This number is now known exactly, because the researcher did the arithmetic.

Step 4: Recognize the statistic is only an estimate. The sample mean x̄ is not the same number as the population mean μ, except by coincidence. If the researcher drew a different sample of the same size from the same population, they would almost certainly compute a different x̄. This sample-to-sample fluctuation is called sampling variation or sampling error (a technical term, not a mistake — it simply reflects that samples are incomplete snapshots of the population).

Step 5: Use the statistic to infer the parameter, with quantified uncertainty. Because sampling variation is itself statistically predictable — larger samples vary less, and the spread of possible sample means follows describable patterns (the foundation of the sampling distribution and later the Central Limit Theorem) — the researcher can report not just x̄ but also a standard error and a confidence interval, which express how far x̄ is likely to be from the true μ. This turns an incomplete measurement into a scientifically honest, quantified estimate of the unknowable truth. The notation convention that encodes all of this is summarized below.

The rule of thumb that keeps this straight for life: Greek letters describe the population truth you cannot see; Roman letters describe the sample number you calculated with your own hands.

The claim that "a sample statistic varies from sample to sample even when the population never changes" is not merely asserted in statistics textbooks — it can be, and has been, demonstrated directly, and it underlies one of the foundational results in the history of statistical inference.

W. S. Gosset and the brewery data (early 1900s). Working as a chemist at the Guinness brewery in Dublin, William Sealy Gosset faced a very practical biological problem: he needed to estimate the true mean quality characteristics (such as yeast cell counts or extract yield) of enormous batches of beer using only small samples, because testing an entire batch destroyed it. Gosset actually performed repeated small samplings from larger, better-characterized batches and observed exactly what the parameter/statistic framework predicts: the mean computed from a small sample of measurements bounced around from sample to sample, never landing exactly on the "true" value calculated from the far larger reference batch. Critically, the amount of bounce (spread) shrank as the sample size grew. Gosset formalized the distribution of how a sample mean, standardized by its sample-based standard error, deviates from the true population mean, publishing under the pseudonym "Student" in 1908 — this became the Student's t-distribution, still the workhorse of small-sample inference in biology today. His work is direct empirical proof that a statistic is a variable quantity clustering around, but not equal to, a fixed parameter.

A simple repeatable classroom/field demonstration. The logic Gosset formalized can be shown directly with any finite, fully-known population — for instance, a jar of 500 dried beans with a known, exact true mean mass μ (because in this special demonstration case, every bean actually was weighed, so the parameter is fully known for teaching purposes). If a class repeatedly draws random samples of n = 10 beans, weighs them, and computes x̄ for each sample, the results always show the same pattern:

• No two samples give exactly the same x̄.
• The x̄ values cluster around the true μ, but individual values scatter above and below it.
• As n increases (say, from 10 beans to 50 beans per sample), the spread of x̄ values across repeated samples shrinks noticeably.

This directly demonstrates, with real measured numbers rather than assertion, that (1) a statistic is a random variable with its own distribution (the sampling distribution), (2) it is unbiased on average — the average of many x̄'s converges toward μ — and (3) its precision improves with sample size, the seed idea for the standard error formula developed later in Standard Error and Confidence Intervals.

What this evidence proves. Together, Gosset's brewery work and repeatable finite-population demonstrations prove that the parameter/statistic distinction is not a notational nicety — it reflects a real, measurable, reproducible statistical phenomenon: samples drawn from the same fixed population predictably differ from one another and from the population truth, in a quantifiable way that shrinks with sample size.

Example 1 — a genetics-flavored population versus sample scenario. Suppose a population geneticist studies a wild population of 12,000 fruit flies fixed for a visible wing-shape locus with two alleles, V (normal) and v (vestigial, recessive). The true population allele frequency of v, call it q, is a parameter — one fixed number describing all 24,000 alleles present at that locus across every fly in the population (diploid, so 2 × 12,000 alleles). No one can genotype every fly, so q is unknown. The researcher captures a random sample of n = 150 flies and genotypes them. Suppose the sample yields:

• 96 flies VV
• 45 flies Vv
• 9 flies vv

The sample allele frequency, the statistic estimating q, is calculated by counting alleles directly observed: Total v alleles = (45 × 1) + (9 × 2) = 45 + 18 = 63 Total alleles sampled = 2 × 150 = 300

q̂ = 63 / 300 = 0.21 This q̂ = 0.21 is a statistic: known exactly, computed from the sample. The true q is a parameter: fixed, but not directly observed. If the researcher recaptured a different 150 flies next week from the same unchanged population, they might compute q̂ = 0.18 or q̂ = 0.24 purely from sampling variation — not because the population's true allele frequency shifted. Only by combining many samples, increasing n, or applying the Hardy-Weinberg Principle with a formal confidence interval can the researcher express how close q̂ = 0.21 is likely to be to the real q.

Example 2 — a physiological/clinical calculation with asymmetric numbers. A hospital wants to know the true mean resting heart rate of all 8,400 patients currently admitted to its cardiology ward across the country's regional hospital network — this true value is a parameter, μ, and it is not practically measurable because patients are scattered across 30 hospitals and heart rate changes minute to minute. A clinical researcher samples n = 55 patients (randomly selected charts from the network's records over one week) and records resting heart rate in beats per minute. The 55 measurements yield:

• Sum of all 55 heart rates = 4,235 bpm
• Sample mean x̄ = 4,235 / 55 = 77.0 bpm
• Sample standard deviation, computed from the same 55 measurements: s = 11.3 bpm

Here x̄ = 77.0 bpm and s = 11.3 bpm are statistics — exact, known numbers computed from the sampled charts. The corresponding population parameters, μ (true mean resting heart rate across all 8,400 patients) and σ (true population standard deviation), remain unknown. Using the statistic and its known sample size, the researcher can compute a standard error of the mean: SE = s / √n = 11.3 / √55 = 11.3 / 7.416 ≈ 1.52 bpm This standard error is not a new parameter — it is a statistic describing how much x̄ itself would be expected to bounce around from sample to sample, and it feeds directly into a 95% confidence interval (roughly x̄ ± 2×SE ≈ 77.0 ± 3.0, i.e., about 74.0 to 80.0 bpm) — a quantified, honest range for the unknown μ, built entirely from sample statistics. This is the conceptual bridge to Standard Error and Confidence Intervals.

The clearest historical illustration of why the parameter/statistic distinction matters comes from William Sealy Gosset, a chemist who needed rigorous small-sample statistics for entirely practical, biologically-flavored industrial reasons. Gosset was born in Canterbury, England, and trained in chemistry and mathematics at Oxford before joining Arthur Guinness & Son brewery in Dublin in 1899. His job required estimating properties of enormous batches of beer and barley — such as the mean starch content of barley varieties or mean yeast viability — from small, cheap samples, because exhaustively testing every unit was destructive and financially absurd for a brewery. Existing statistical theory at the time assumed sample sizes large enough that the normal distribution was an adequate approximation for the sampling behavior of the mean; Gosset's brewery samples were often tiny (n < 10), and he noticed the standard large-sample formulas gave systematically overconfident, too-narrow estimates of uncertainty when n was small. He derived the correct small-sample distribution for the standardized deviation of a sample mean from the true population mean — what is now called the t-distribution — and published it in Biometrika in 1908 under the pseudonym "Student," because Guinness had a policy discouraging its scientists from publishing under their own names (to protect trade secrets). Gosset's work, later extended and popularized by Ronald A. Fisher, remains the mathematical backbone of small-sample inference across all of experimental biology.

The parameter/statistic distinction is not a historical curiosity — it is invoked, explicitly or implicitly, every time a biological or medical number is reported with any honesty about its uncertainty.

• Clinical trials and drug approval: A trial reports a sample statistic — the mean reduction in blood pressure among enrolled patients — but regulatory decisions are actually about the unknown population parameter, the true effect in all future patients who might take the drug. Confidence intervals and p-values exist specifically to communicate how far the sample statistic might be from that unknown truth.
• Public health surveillance: Estimated disease prevalence (e.g., "12% of adults in this region have condition X") is always a statistic from a surveyed sample, never a census of every person; agencies like the CDC and WHO explicitly report confidence intervals around such statistics because the true population prevalence parameter is never directly known.
• Conservation biology and population viability analysis: Estimating a species' true population size, genetic diversity, or extinction risk parameter from mark-recapture samples or genetic subsampling requires the same logic — a handful of tagged animals recaptured gives a statistic used to infer an unknowable true population parameter, directly informing endangered species policy.
• Agricultural breeding programs: Crop yield trials measure sample statistics (mean yield across sampled plots of a new cultivar) to estimate the population parameter (true mean yield across all future plantings of that cultivar), guiding which varieties get released to farmers.
• Biotechnology quality control: Pharmaceutical manufacturing tests a sample of vials or doses from each production lot (a statistic) to make inferences about the entire lot's true potency or contamination rate (a parameter), directly analogous to Gosset's original brewery problem.

Ask the tutor: If my sample size is very large — say, n = 10,000 — is it ever acceptable to treat the sample statistic as if it were the true population parameter?