A relative frequency is the proportion (or percentage) of a dataset’s total observations that fall into a given category or class: relative frequency = count / total. A frequency distribution is the full table or chart showing how observations are spread across all categories or class intervals. Together, these two ideas are the foundation of descriptive statistics for both categorical and continuous variables, and relative frequency in particular does double duty as an empirical estimate of probability — the bridge between counting data and reasoning about chance.
Frequency, Relative Frequency, Cumulative Frequency, and Relative Cumulative Frequency
Four related quantities are easy to confuse because they’re often built from the same table:
- Frequency (f) — the raw count of observations in a category or class interval.
- Relative frequency (f / n) — the frequency divided by the total number of observations (n), expressed as a proportion (0 to 1) or a percentage. This is what makes categories or classes of different sizes, or datasets of different total size, comparable.
- Cumulative frequency — the running total of frequencies up to and including a given category or class, in order.
- Relative cumulative frequency — the cumulative frequency divided by n; equivalently, the running total of relative frequencies. This is the value that tells you what proportion of the data falls at or below a given point, and it is the discrete-data building block of the cumulative distribution function (CDF).
Illustrative example, not real data. Suppose a research-integrity office reviews 40 submitted manuscripts and records the number of data-entry errors flagged in each one during a pre-submission check:
| Errors found | Frequency (f) | Relative frequency (f/n) | Cumulative frequency | Relative cumulative frequency |
|---|---|---|---|---|
| 0 | 8 | 0.20 (20%) | 8 | 0.20 (20%) |
| 1 | 14 | 0.35 (35%) | 22 | 0.55 (55%) |
| 2 | 10 | 0.25 (25%) | 32 | 0.80 (80%) |
| 3 | 5 | 0.125 (12.5%) | 37 | 0.925 (92.5%) |
| 4 or more | 3 | 0.075 (7.5%) | 40 | 1.00 (100%) |
| Total | n = 40 | 1.00 (100%) | — | — |
Two checks confirm the table is internally consistent, and are worth running on any frequency distribution you build: the frequency column sums to n, and the relative frequency column sums to 1.00 (or 100%, allowing for rounding). The cumulative frequency in the last row always equals n; the relative cumulative frequency in the last row always equals 1.00.
Why Relative Frequency, Not Just Frequency
Raw frequency is only interpretable relative to n. A frequency of 14 means something different in a sample of 40 than in a sample of 4,000. Relative frequency strips that dependency out, which is what makes it possible to:
- Compare the shape of two distributions built from samples of different sizes (a class of 22 students against a cohort of 2,200 survey respondents).
- Compare a subgroup against the whole (what proportion of early-career respondents used a data management plan, versus the sample overall) — see descriptive statistics for how relative frequency fits alongside measures of central tendency and dispersion.
- Overlay two frequency distributions on the same chart without one simply dwarfing the other because it has more observations.
Whether to report a proportion (0 to 1) or a percentage (0 to 100) is a style choice; percentages are usually easier to read in prose and tables, proportions are usually what’s used inside further calculations (chi-square, confidence intervals). Report percentages to a sensible number of decimal places for the sample size — more precision than the data can support is misleading, not rigorous.
Relative Frequency as an Empirical Estimate of Probability
This is the conceptual payoff of the whole topic. If you flip a fair coin and record the relative frequency of heads, you don’t expect exactly 0.50 after 10 flips — you might see 0.40 or 0.70. But as the number of flips grows, the relative frequency of heads gets closer and closer to the true underlying probability, 0.50. This is the law of large numbers: as the number of independent trials (n) increases, the relative frequency of an event converges toward its true probability.
This is the practical justification for treating an observed relative frequency as an estimate of probability at all — the entire frequentist interpretation of probability rests on it. It is also why small-n relative frequencies are volatile and untrustworthy as probability estimates: with only a handful of trials, the relative frequency can sit far from the true probability purely by chance, and it takes a much larger n than intuition suggests before that gap reliably narrows. This has direct, practical consequences for research design and reporting:
- An observed proportion from a small pilot sample (a response rate, an adverse-event rate, an assay failure rate) is a noisy estimate of the true underlying probability, not a precise measurement of it — report a confidence interval alongside it, not just the point estimate, whenever the sample is small.
- Statistical power calculations exist precisely because of this convergence behavior: a study needs enough observations for the observed relative frequency (or difference in relative frequencies between groups) to be a trustworthy estimate of the true effect, not noise.
- The law of large numbers describes convergence in the long run; it says nothing about any individual trial or short run — a common misreading (the “gambler’s fallacy”) wrongly expects short-run outcomes to self-correct trial-by-trial to match the long-run probability.
Building a Frequency Distribution for Continuous Data: Class Intervals and Bin Width
Categorical and discrete data (like the manuscript-errors example above) sort naturally into distinct values. Continuous data — time, concentration, distance, reaction time — has to be grouped into class intervals (also called bins) before a frequency distribution can be built at all, because in principle no two continuous measurements are ever exactly equal.
Illustrative example, not real data. Turnaround time, in days, from submission to first decision for 30 IRB protocol reviews, grouped into class intervals of width 5:
| Turnaround time (days) | Frequency (f) | Relative frequency (f/n) |
|---|---|---|
| 0–4 | 3 | 0.10 (10%) |
| 5–9 | 9 | 0.30 (30%) |
| 10–14 | 11 | 0.367 (36.7%) |
| 15–19 | 5 | 0.167 (16.7%) |
| 20–24 | 2 | 0.067 (6.7%) |
| Total | n = 30 | 1.00 (100%) |
Choosing class-interval (bin) width is a genuine trade-off, not a mechanical step:
- Too few, wide intervals smooth away real structure — a bimodal distribution can look unimodal if the bins are wide enough to merge the two peaks.
- Too many, narrow intervals produce a table (or histogram) so ragged with sampling noise that the underlying shape is hard to see, and some intervals may have very few or zero observations.
- Common starting rules of thumb (Sturges’ rule, the square-root rule, Scott’s or Freedman-Diaconis’ rules based on the data’s spread) give a reasonable starting bin count for a given n, but none is universally correct — the right choice depends on what structure in the data you need to see, and it’s reasonable to try a few widths and compare.
A frequency distribution for continuous data is the tabular form of a histogram — the histogram is simply that same relative-frequency-by-class-interval table drawn as bars, which is why bin-width choice matters identically for both. Once you have a distribution built, its shape (symmetric, left-skewed, right-skewed) is usually described using the vocabulary in skewness: left-skewed vs. right-skewed distributions.
Relative Cumulative Frequency, the Empirical CDF, and Percentiles
Running the relative-frequency column as a cumulative total (as shown in the first worked table) answers a different question than the frequency table alone: not “how many observations fall in this class?” but “what proportion of all observations fall at or below this point?” That running total is the discrete, sample-based version of the empirical cumulative distribution function — a step function that jumps up by each class’s relative frequency as you move through the ordered classes.
Relative cumulative frequency is also the direct route to percentiles and quartiles: the value of x where the relative cumulative frequency first reaches (or crosses) 0.25, 0.50, or 0.75 is, respectively, an estimate of the 25th percentile (first quartile), median, and 75th percentile (third quartile) of the data. See interquartile range & five-number summary for how those percentile-based summaries are calculated and used from grouped data of exactly this kind.
Relative Frequency in Categorical Data: Row, Column, and Total Percentages
When two categorical variables are cross-tabulated into a contingency table, “relative frequency” stops being a single unambiguous number — there are three different denominators you could divide by, and choosing the wrong one is one of the most common misreadings of a cross-tab.
Illustrative example, not real data. 120 survey respondents cross-tabulated by career stage and whether they reported using a data management plan (DMP) on their current project:
| Career stage | Used a DMP | Did not use a DMP | Row total |
|---|---|---|---|
| Early-career | 18 | 42 | 60 |
| Mid/senior-career | 36 | 24 | 60 |
| Column total | 54 | 66 | 120 |
- Row percentage divides by the row total: among early-career respondents, 18/60 = 30% used a DMP. This answers “within this group, what proportion did X?”
- Column percentage divides by the column total: among respondents who used a DMP, 18/54 = 33.3% were early-career. This answers a different question — “within this outcome, what’s the composition by group?” — and is easy to mistake for the row percentage if the denominator isn’t stated.
- Total (or cell) percentage divides by the grand total: 18/120 = 15% of all respondents were both early-career and DMP-using. This is the least commonly useful of the three for comparing groups, because it conflates group size with rate.
Reporting a cell’s percentage without saying which denominator it’s relative to is genuinely ambiguous, not just imprecise — “30% of early-career respondents” (row) and “33.3% of DMP-users” (column) are both true, correct statements about the same cell, and they answer different questions. This exact table is also the input to a chi-square test of independence, which formally tests whether the row and column variables are associated by comparing observed cell frequencies against the frequencies expected if the two variables were unrelated — get the row/column percentages right first, because they’re usually what makes a significant chi-square result interpretable in plain language afterward.
Reporting Relative Frequency Correctly: Always Give n
A percentage without its denominator is not fully interpretable. “33% of respondents reported X” means something very different, and carries very different statistical weight, depending on whether it’s 1 of 3, 10 of 30, or 1,000 of 3,000 — all three are 33%. Small-n percentages are the most common way this goes wrong in practice: 1 of 3 (33%) and 1 of 4 (25%) sound like meaningfully different rates, but with samples that small the difference is almost certainly noise, and a single additional observation swings the percentage by 8–33 points.
Good practice, consistent with reporting conventions used across the health and social sciences:
- Always state n alongside any percentage, either inline (“18 of 60, 30%”) or in a table footnote/column.
- Be cautious reporting percentages at all for very small subgroups (single digits); consider reporting the raw count instead, or a count with a confidence interval.
- Keep the denominator consistent and explicit when comparing percentages across groups of different sizes — and, per the contingency-table section above, always state which denominator (row, column, or total) a percentage is relative to.
- Don’t over-report decimal precision a small n can’t support — 12.5% from 5/40 is defensible; 12.50000% is not.
Frequently Asked Questions
What is the difference between frequency and relative frequency?
Frequency is the raw count of observations in a category or class. Relative frequency is that count divided by the total number of observations (n), expressed as a proportion or percentage. Relative frequency is what makes categories, classes, or entire samples of different sizes comparable.
How do you calculate relative frequency?
Divide the frequency of a category or class by the total number of observations: relative frequency = f / n. Multiply by 100 to express it as a percentage instead of a proportion.
What is a relative frequency distribution?
A table or chart showing the relative frequency of every category or class interval in a dataset, rather than the raw counts. The relative frequencies across all categories or classes should sum to 1.00 (or 100%).
Is relative frequency the same as probability?
Not exactly, but they’re closely linked. Relative frequency is an observed proportion from actual data; probability is the underlying, usually unknown, true likelihood of an event. Relative frequency is used as an empirical estimate of probability, and by the law of large numbers it converges toward the true probability as the number of observations grows — but any single sample’s relative frequency, especially from a small sample, can differ from the true probability.
How many class intervals should a frequency distribution have?
There’s no single correct number — it’s a trade-off between smoothing away real structure (too few, wide intervals) and drowning the shape in sampling noise (too many, narrow intervals). Rules of thumb (Sturges’ rule, the square-root rule, Scott’s or Freedman-Diaconis’ rules) give a reasonable starting point based on sample size or data spread, but comparing a couple of different bin widths on the same data is standard practice.
What is relative cumulative frequency used for?
It shows what proportion of a dataset falls at or below a given value, which is the basis for reading off percentiles, quartiles, and the median from grouped data, and is the sample-based version of the cumulative distribution function.
Why should you report the sample size (n) alongside a percentage?
Because a percentage without its denominator is not fully interpretable — the same percentage can represent very different amounts of evidence depending on n, and percentages from small samples (single digits or low tens) are especially volatile and can be misleading if presented without the underlying count.
