Can an observed frequency be zero?

Yes. No recorded observations may fall in a category, even when that outcome is possible.

Is an expected frequency necessarily a whole number?

No. It is a model-based expected count and can be fractional.

Are observed frequencies probabilities?

No. Dividing by an appropriate total gives relative frequencies, which may estimate probabilities under suitable sampling conditions.

AI Glossary / Statistics and probability / Beginner

Observed frequencies

Definition

Observed frequencies are the counts actually recorded in categories, intervals, or cells of a frequency table. They describe how many observations fell into each specified group in the collected dataset. They are distinguished from expected frequencies, which are calculated from a model or hypothesis. An observed frequency records what the data show under the chosen counting rules; it does not by itself establish the population probability or explain why the pattern occurred.

What it means

If a survey records 18 people choosing one option, 18 is that option's observed frequency. Dividing by the relevant total produces a relative frequency, such as a proportion or percentage. The count and the proportion convey related but different information.

The observational unit and denominator matter. A count of support messages is not necessarily a count of customers, because one customer may send several messages. Categories and inclusion rules determine what each cell means, so they should be documented before interpreting differences across cells or datasets.

How it works

Define the population or sample being summarized and decide what constitutes one observation. Establish categories that suit the question. For a simple one-category-per-observation table, categories should be mutually exclusive and collectively cover the observations included in the denominator.

Count each eligible observation in its correct cell, handling missing, unknown, or out-of-scope values consistently. Check that the cell counts add to the intended total. Calculate relative frequencies only after choosing a clear denominator.

If comparing observed and expected frequencies, derive expectations from the specified hypothesis. For example, a goodness-of-fit analysis multiplies the sample size by hypothesized category probabilities. A contingency-table analysis of independence uses a different construction based on margins. Inferential procedures also require appropriate assumptions about sampling, independence, and the adequacy of expected cell sizes; the existence of a table does not automatically justify a test.

Three practical examples

An AI evaluation counts answers classified as correct, partly correct, or incorrect under a predefined rubric.
A customer survey reports how many respondents chose each preferred support channel.
A quality team records defects by type and compares the observed pattern with a specified reference distribution.

Think of it this way

Think of sorting a collection of buttons into labeled containers and counting each container. Those counts are observed frequencies. A supplier's claim about the usual mix would produce expected counts for comparison, but it would not replace the buttons actually counted.

Worked example

Suppose 40 fictional support requests are classified into three mutually exclusive categories: billing, access, and other. The observed frequencies are 18, 12, and 10. They sum to 40, and their relative frequencies are 45%, 30%, and 25%.

A prespecified reference model assigns probabilities of 0.50, 0.30, and 0.20. Its expected frequencies for 40 requests are 20, 12, and 8. The differences between observed and expected counts are therefore −2, 0, and 2.

For illustration, the Pearson goodness-of-fit statistic is (18 − 20)²/20 + (12 − 12)²/12 + (10 − 8)²/8 = 0.7. This calculation summarizes discrepancy; its interpretation requires the model and sampling assumptions. It does not mean that the observed counts should be changed to match the expected ones, or that the reference probabilities are proved correct.

Illustrative case study

Fictional case study: Cedar Support publishes a dashboard of AI-assisted service outcomes. Initially, the chart counts every message as if it represented a separate request. Long conversations consequently contribute more observations than short ones.

The team defines one completed request as the unit, specifies a consistent outcome rubric, and separates unclassified requests from classified ones. It then reports both counts and percentages with their denominators. When comparisons use a reference model, expected counts appear in a separate column. This fictional case illustrates transparent counting and does not claim a real change in service performance.

Common misconceptions and mistakes

Common mistakes include calling percentages frequencies without clarifying the scale, confusing observed counts with expected counts, and omitting categories that have zero observations. A zero observed count does not prove that an event is impossible.

Overlapping categories can make totals exceed the number of observations, which may be appropriate for multiple-response questions but must be explained. Weighted survey totals can also be fractional and should not be described as raw respondent counts. Finally, repeated observations from the same person may be dependent; treating them as independent can distort uncertainty even when the counting itself is correct.

Business use

Observed frequencies provide the foundation for operational dashboards, survey summaries, error analyses, and model evaluation. Counts help businesses see the volume behind a percentage: a 50% rate based on two cases carries different information from the same rate based on thousands.

AI-generated reports should retain units, time windows, inclusion rules, and missing-data treatment. Validation can check totals and detect duplicated records, but human judgment may still be needed to determine whether the categories answer the intended question. Clear counting rules make comparisons reproducible.

Check your understanding

A table records 15, 9, and 6 observations across three exclusive categories. What is the relative frequency of the second category?

Answer: The total is 30, so its relative frequency is 9/30 = 0.30, or 30%. Its observed frequency remains nine.

Related terms

Relative frequency
Expected frequencies
Contingency table
Categorical data
Frequency table
expected frequency
goodness-of-fit test
chi-square statistic
denominator
sampling variation

Learn next

Practice building frequency tables from raw records and checking their totals. Then study expected counts under goodness-of-fit and independence models. Learn how sparse cells, survey weights, and repeated observations affect the choice and interpretation of inferential methods.

Goodness-of-fit

Sources

David Howell's Statistical Methods for Psychology glossary distinguishes observed from expected cell frequencies: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html

The NIST Engineering Statistics Handbook explains their use in the chi-square goodness-of-fit statistic: https://www.itl.nist.gov/div898/handbook/eda/section3/eda35f.htm

The counts, calculation, and fictional case are original teaching examples.

Previous: Observational study Next: Observer-expectancy effect