AI Glossary / Statistics and probability / Beginner

Reference group

Definition

A reference group is a group of observations, or a defined group that could be observed, used as a point of comparison in a study. It supplies a baseline for describing differences, ratios or model coefficients. NIST distinguishes the term from a deliberately constructed randomized control group: calling a group a reference does not imply that its membership was assigned randomly. The group must be defined clearly enough for readers to understand what the comparison actually means.

What it means

Statements such as higher, lower or twice as frequent require a comparison. The reference group supplies that context. It may be an existing population category, a historical group, a designated baseline in a model or a suitable experimental comparator. Those choices have different evidential implications. A numerical contrast with a reference group can be descriptive without establishing that group membership caused the difference. The suitability of the reference depends on the study question and how the observations were obtained.

How it works

Specify the target comparison and define the reference group's inclusion criteria, time period and measurement process. Check whether the groups are comparable on factors relevant to the outcome. Choose an effect measure, such as a difference, ratio or model-based contrast, and state its direction. In regression with categorical predictors, a reference category can be represented through coding choices. Interpret coefficients relative to that category and include appropriate uncertainty. Changing the reference changes the representation of comparisons, not the underlying observations.

Three practical examples

Example 1: A service report compares a new customer segment with an established segment defined as the reference. Example 2: A study compares current measurements with a historical reference population while acknowledging changes over time. Example 3: A regression model uses one product category as the baseline for category coefficients. In each case, the reference should serve the question rather than be selected because it makes another group appear especially favorable.

Think of it this way

Think of measuring height from a clearly marked starting level. Moving the starting level changes the reported offsets but does not move the objects themselves. Similarly, choosing a different reference category changes the way differences are expressed. The analogy has a limit: selecting a genuinely different comparison population can also change the substantive question, so the reference is not always a purely cosmetic choice.

Worked example

Suppose group A has 20 events among 100 observations and group B has 30 among 100. Using A as reference, B's risk difference is 0.30 - 0.20 = 0.10, or ten percentage points, and its risk ratio is 0.30/0.20 = 1.50. Using B as reference, A's risk difference is -0.10 and its risk ratio is 0.20/0.30 = 2/3, approximately 0.667. The data have not changed. In a separate simple mean model, if A's mean is four and B's is seven, coding A as reference gives intercept four and B coefficient three. Coding B as reference gives intercept seven and A coefficient minus three. These descriptive transformations do not establish a causal effect of belonging to either group.

Illustrative case study

Fictional case study: Linden Services publishes a comparison of response rates using its smallest segment as the reference. Reviewers find that this makes ratios unstable and does not match the business question. The team explains the choice, reports absolute rates and uncertainty, and adds the more relevant comparison. This fictional case illustrates transparent reporting, not a rule that small groups must never be references or that one baseline is universally correct.

Common misconceptions and mistakes

Common mistakes include assuming a reference group was randomized, treating a baseline category as inherently normal or desirable, and omitting the reference when reporting a coefficient. Ratios reverse by reciprocation rather than by simply changing the sign. Historical references can differ in measurement, selection and background conditions. Adjustment for observed covariates does not guarantee removal of all confounding. A statistically convenient reference may also be substantively misleading if it obscures the comparison readers need.

Business use

Business use: Reference groups support benchmarking, segmentation and model interpretation. Define them in dashboards and reports so readers can interpret every contrast. Present absolute outcomes alongside relative measures, especially when the reference rate is small. Explain whether the comparison is descriptive, adjusted or supported by random allocation. Review reference choices when populations, products or measurement methods change. Avoid language that turns a coding convention into an unsupported judgment about which group is standard or superior.

Check your understanding

Why does the risk ratio change from 1.50 to about 0.667 when the reference reverses?

Answer: The numerator and denominator switch, so the ratios are reciprocals.

Does a higher observed rate in B show that B membership caused events?

Answer: No. The comparison alone does not address confounding, selection or other causal assumptions.

Related terms

while risk ratios compare their ratios

Learn next

Calculate the same difference and ratio using each group as the reference, and explain how interpretation changes. Recode a simple categorical regression and verify that predictions remain the same. Then compare a concurrent randomized comparator with a historical reference group to understand why identical arithmetic can carry different evidential strength.

Sources

Sources: NIST/SEMATECH e-Handbook glossary, reference group: https://www.itl.nist.gov/div898/handbook/glossary.htm ; Penn State STAT 501, Regression Methods: https://online.stat.psu.edu/stat501/index . These institutional sources support comparison-group and regression interpretation. The reversed-reference calculations and fictional service-report case are original illustrations, with descriptive comparisons explicitly separated from causal claims.

Previous: Reference counts Next: Reference metadata