Types of Biological Data

Biological data is any recorded observation about living systems — a fish's length, a patient's diagnosis, the number of eggs in a nest. All such data can be classified along one primary axis: categorical versus numerical.

Categorical data place each observation into one of a limited set of groups or labels rather than a numeric quantity. Categorical data are further divided into two types. Nominal data are categories with no inherent order — blood type (A, B, AB, O), species identity, sex, habitat type. There is no sense in which "A" is greater than "O." Ordinal data are categories that do have a natural order, but the intervals between categories are not necessarily equal or even meaningful — tumor stage (I, II, III, IV), a Likert-scale pain rating (1–5), or dominance rank in a primate troop. Stage III is worse than Stage II, but we cannot say the "distance" from I to II equals the distance from III to IV.

Numerical data (also called quantitative or measurement data) record an actual quantity, and arithmetic operations like addition and averaging are meaningful. Numerical data split into discrete and continuous subtypes. Discrete data can take only specific, usually whole-number, values, typically because they arise from counting — the number of offspring in a litter, the number of chromosomes, the number of bacterial colonies on a plate. There is no such thing as 3.7 chromosomes. Continuous data, by contrast, can in principle take any value within a range, limited only by the precision of the measuring instrument — height, blood pressure, enzyme concentration, body temperature. Between any two continuous values, another value is always conceivable. This nominal/ordinal/discrete/continuous scheme is the biological instantiation of a more general statistical framework called <a class="wiki-link" href="/wiki/bio-biostat-variables-and-measurement-scales" data-track="wiki_article.body.wiki_link">Variables and Measurement Scales</a> (nominal, ordinal, interval, ratio scales), which refines these categories further by asking whether zero is a true "nothing" and whether ratios are meaningful. For now, the categorical/numerical split — and discrete/continuous within numerical — is the essential first sort.

Step 1: Identify what is actually being recorded. For each variable, ask: "Is this a label, or a quantity?" A veterinarian recording "breed" of dog is recording a label (categorical, nominal). A veterinarian recording "weight in kg" is recording a quantity (numerical).

Step 2: If categorical, ask whether the categories have a natural order. "Fur color" (black, brown, white) has no natural order — nominal. "Disease severity" (mild, moderate, severe) has a clear order — ordinal. This distinction matters because ordinal data permit median and rank-based statistics, while pure nominal data are limited to counts, proportions, and modes.

Step 3: If numerical, ask whether the value arises from counting or measuring. Counting gives discrete data: the number of petals on a flower, the number of surviving offspring, colony-forming units on an agar plate. These values are restricted to a countable set (usually non-negative integers) and jumps between values are indivisible — you cannot have 4.5 petals. Measuring gives continuous data: length, mass, concentration, pressure, time. Any two continuous measurements, however close, can in principle be separated by a value even closer, subject only to the resolution of the ruler, scale, or assay.

Step 4: Recognize the practical blur at high counts. When discrete counts get very large — for instance, red blood cell counts per microliter of blood, often in the millions — the individual whole-number steps become statistically indistinguishable from a continuous scale, and such data are often treated as continuous for modeling purposes. The underlying biology, though, is still discrete: you cannot have half a red blood cell. This is a pragmatic simplification, not a change in what the variable actually is.

Step 5: Let data type dictate summary statistics and graphing choices. This is the payoff. For nominal data, only counts, proportions, and the mode are meaningful — you tally how many organisms fall in each category and display them with a bar chart or pie chart. For ordinal data, the median and percentiles are meaningful (because categories can be ranked), but a mean is often statistically dubious because the "distance" between ranks is undefined. For discrete numerical data, mean, median, and standard deviation are all valid, though the mean itself may come out as a non-whole number (e.g., a mean litter size of 4.3 pups is a legitimate summary even though no single litter has 4.3 pups). For continuous numerical data, the full arsenal of descriptive and inferential statistics applies, and the natural graphical display is a histogram, not a bar chart, because there are no gaps between adjacent values.

The visual difference in Figure 2 is not cosmetic. Discrete bar charts (sometimes drawn with gaps, or with bars centered on integers) reflect the fact that non-integer clutch sizes are biologically impossible; continuous histograms have touching bars because any blood-pressure reading between bins is conceivable — the bin boundaries are an artifact of how finely we chose to group the data, not a property of blood pressure itself.

Unlike phenomena such as inheritance or natural selection, "types of biological data" is not a hypothesis tested by a single classic experiment — it is a measurement framework. But the necessity of getting this classification right is itself demonstrated empirically, most famously by the history of how biologists have handled ordinal data mistaken for continuous data, and by foundational work in biometry that depended on correctly separating discrete counts from continuous measurements.

Karl Pearson and Walter Weldon's biometric surveys (1890s). When Pearson and Weldon set out to quantify variation in natural populations — crab carapace widths, shrimp segment counts — they had to explicitly decide, for each trait, whether it was a continuously varying measurement (like carapace width, measured with calipers) or a discrete meristic count (like the number of abdominal segments or fin rays). Weldon's crab data (collected at the Naples marine station) treated frontal breadth as continuous, fitting it to the normal curve; counts of gill rakers or vertebrae in fish, by contrast, were tabulated as discrete frequency distributions. This foundational distinction — continuous morphometric traits versus discrete meristic traits — became a standard organizing principle in biometry and remains embedded in modern quantitative genetics, where "meristic" (discrete, countable) traits like litter size or bristle number are still modeled differently from continuous traits like body mass.

R.A. Fisher and the chi-square test for categorical data (1920s). Fisher's development and popularization of the chi-square goodness-of-fit and contingency-table tests depended entirely on recognizing that some data — phenotype categories in a genetic cross, for instance — are inherently categorical counts, not measurements, and therefore cannot be analyzed with means and t-tests. Fisher's famous 1936 re-examination of Gregor Mendel's pea data used chi-square statistics precisely because Mendel's results (round vs. wrinkled seeds, yellow vs. green cotyledons) were categorical counts of nominal categories — the correct test had to match the correct data type, and Fisher's analysis (controversially) suggested Mendel's category counts fit expected ratios too well, a case still debated by statisticians and historians of science.

Ronald Fisher and Karl Pearson also clashed over exactly this kind of classification question — how many "degrees of freedom" a contingency table has depends on correctly identifying how many independent categories exist, a dispute settled by Fisher's 1922 correction to Pearson's original chi-square formula. The resolution of that dispute is itself evidence that rigorous handling of categorical (nominal) data, as distinct from continuous data, is not a pedantic footnote but a matter that determined the validity of decades of published genetics results. Together, these episodes show that the categorical/numerical, discrete/continuous framework is not an arbitrary taxonomy invented for textbooks — it emerged because biologists doing real quantitative work (crab measurements, pea-plant ratios, contingency tables) kept discovering that the wrong statistical tool, applied to the wrong data type, produced wrong conclusions, and the field had to formalize the distinction to make progress.

Example 1 — A mixed dataset from a fisheries survey. A marine biologist surveys 200 individual fish caught in a trawl net and records, for each fish: (1) species (cod, haddock, or whiting), (2) maturity stage (immature, maturing, mature, spawning — an increasing severity/maturity scale), (3) number of parasitic copepods found on the gills, and (4) body length in centimeters measured with digital calipers to 0.01 cm precision. Classify each variable:

Suppose the parasite counts across 200 fish give a mean of 3.4 copepods per fish. This is a legitimate summary statistic even though no individual fish can carry 3.4 copepods — the individual data are discrete, but the summary (a mean) need not itself be a whole number. If a colleague instead reported "the median maturity stage was 2.5," that would be more awkward: since maturity stage is ordinal, not truly numerical, a value like 2.5 lacks the same biological interpretability, even though the arithmetic of computing it is identical to computing a numerical median. This is exactly why ordinal data are often better summarized by the actual category ("maturing") nearest the middle, or by the proportion in each named category, than by an interpolated numeric value.

Example 2 — Clinical blood pressure classification with asymmetric group sizes. A clinic screens 350 adults and classifies systolic blood pressure (a continuous variable, measured in mmHg) into four ordinal categories used in hypertension staging:

Two questions probe the data-type distinction directly. (a) What is the underlying data type of "systolic blood pressure," and what is the type after binning into these four labels? The raw measurement (e.g., 128.4 mmHg) is continuous numerical data — it can, in principle, take any value, limited by the sphygmomanometer's precision. Once binned into "Normal / Elevated / Stage 1 / Stage 2," the same information becomes ordinal categorical data: the categories have a clear order (worsening hypertension) but the underlying continuous scale has been discarded. This is a common and sometimes costly move in clinical biostatistics — binning throws away information (two patients at 121 mmHg and 129 mmHg are both merely "Elevated," even though they differ by 8 mmHg) purely to gain interpretability for clinical decision-making. (b) Using the binned data, what proportion of patients have at least Stage 1 hypertension, and is a "mean category" meaningful? Adding Stage 1 and Stage 2 patients:

Forty percent of patients have at least Stage 1 hypertension — a perfectly valid statistic because it is a proportion on categorical (here, ordinal) data. But if someone tried to code Normal=1, Elevated=2, Stage 1=3, Stage 2=4, and compute a "mean category" of, say, 2.13, that number is statistically dubious: it assumes the "distance" from Normal to Elevated equals the distance from Stage 1 to Stage 2, which is not guaranteed by the (unequal, clinically defined) mmHg ranges underlying each label. The honest analysis either reports proportions in each ordinal category or returns to the original continuous mmHg values and computes a genuine numerical mean (which, from the original unbinned data, might be reported as "mean systolic pressure = 127.3 mmHg, SD = 14.6 mmHg").

The modern formalization of measurement scales — nominal, ordinal, interval, ratio, which underpins the categorical/numerical split used throughout biology — is usually credited to the psychologist-turned-psychophysicist Stanley Smith Stevens, whose 1946 paper "On the Theory of Scales of Measurement" (published in Science) gave biologists and statisticians the vocabulary still used today. Stevens was not primarily a biologist; he trained as a psychologist at Harvard, working under E.G. Boring, and spent much of his career studying the perception of loudness and sensory thresholds — work that forced him to grapple constantly with the question, "what kind of number is this, and what can I legitimately do with it?" That question, born from psychoacoustics, turned out to be exactly the question biologists needed answered for everything from Mendelian phenotype categories to blood pressure readings. Stevens's framework built on, and helped organize retrospectively, decades of earlier biometric practice — the Pearson–Weldon distinction between meristic (discrete count) and continuous morphometric traits, and Fisher's insistence on matching statistical tests to categorical versus numerical data. Stevens gave a name and a hierarchy to distinctions biologists and statisticians had been making empirically, often clumsily, for half a century.

Clinical diagnostics. Electronic health records mix all four data types constantly: blood type (nominal), cancer stage (ordinal), white blood cell count (discrete), and blood pressure or cholesterol level (continuous). Choosing the wrong statistical test — say, a t-test on cancer stage instead of a rank-based test — can produce misleading trial results, which is why clinical biostatisticians insist on pre-specifying data types before analysis.

Genomics and variant calling. Genotype at a single nucleotide polymorphism is categorical (AA, Aa, aa — nominal, though sometimes treated as ordinal "additive dosage" of the risk allele: 0, 1, 2 copies). Correctly modeling this dosage-coded genotype as an ordinal/discrete numeric variable, rather than a pure nominal category, is central to modern genome-wide association studies (GWAS).

Conservation ecology and species counts. Population census data (number of individuals of an endangered species observed per transect) are discrete counts, often modeled with Poisson or negative-binomial distributions specifically because they are non-negative integers, not continuous quantities — using ordinary linear regression (built for continuous, potentially negative values) on raw count data can produce nonsensical negative population predictions.

Agricultural breeding programs. Traits selected in crop and livestock breeding split cleanly along this framework: continuous traits (milk yield in liters, kernel weight in grams) are analyzed with quantitative genetic models assuming a continuous distribution, while discrete meristic traits (number of piglets per litter, number of kernels per ear) require count-based models. Misapplying continuous-trait breeding-value models to a discrete trait can bias estimated heritability.

Public health surveillance. Categorizing continuous risk factors (BMI, blood glucose) into ordinal clinical categories (underweight/normal/overweight/obese; normal/prediabetic/diabetic) is a routine but consequential simplification, trading statistical power for interpretability — understanding this trade-off is essential for correctly designing and interpreting public-health screening programs. Every one of these applications depends on the same first step introduced here: correctly sorting the data before choosing the tool, a skill that directly sets up Variables and Measurement Scales and Frequency Distributions, where these categories are turned into tables, graphs, and eventually statistical models.

Ask the tutor: How do I decide whether to treat a borderline variable — like age in years — as discrete or continuous in my own dataset?