Can an expected count exceed the observed count?

Yes. The model and the sample need not match exactly.

Must expected frequencies be whole numbers?

No. Expectations can be fractional.

Are expected counts always fixed before collecting data?

No. Independence-table expected counts are commonly fitted using observed margins.

AI Glossary / Statistics and probability / Intermediate

Expected frequencies

Definition

Expected frequencies are the counts predicted for categories or table cells by a specified probability model, often the null hypothesis in a statistical test. They represent model-based averages over hypothetical repetitions, not necessarily the counts that will appear in a particular sample. Expected frequencies can be fractional even though observed counts of people or events are integers.

What it means

Suppose a model assigns probability 0.3 to a category in a sample of 100 independent observations. Its expected frequency is 30. The observed frequency could be 26, 34, or another value because sampling varies.
The word “expected” is mathematical, not a statement about what a manager hopes to see. An expected count is meaningful only in relation to a model, sample size, and definition of the categories. Changing the null hypothesis can change the expected frequencies.

How it works

For a goodness-of-fit problem with fixed sample size n and specified category probabilities p_i, calculate E_i = np_i. The probabilities must sum to one across the mutually exclusive, exhaustive categories, so the expected counts sum to n.
For a two-way contingency table under independence, estimate cell probabilities from the observed marginal proportions. The expected count in row i and column j is E_ij = R_i C_j/n, where R_i and C_j are the row and column totals. These fitted expected counts reproduce the observed margins.
Compare observed and expected counts using an appropriate statistic. Pearson's chi-square statistic sums (O − E)²/E across cells with positive expected counts. Its reference distribution depends on the design and fitted parameters. For an ordinary r-by-c independence test without special constraints, the usual degrees of freedom are (r − 1)(c − 1).
The familiar chi-square reference is an approximation. Sparse expected counts can make it unreliable, and there is no single universal count threshold appropriate to every setting. Consider a suitable exact or simulation-based method when warranted by the table and sampling design.
Distinguish an observed zero from a structural zero. An observed zero means no cases happened to enter a possible cell. A structural zero means the combination is impossible by design or definition; it changes the model and potentially the degrees of freedom.
If probabilities are estimated from the same data, account for that estimation when calibrating a goodness-of-fit test. Weighted surveys, clustered observations, and repeated measurements also require methods consistent with their design. Treating weighted totals as ordinary independent counts can produce misleading uncertainty.

Three practical examples

A service analyst compares request types with a prespecified historical probability distribution using expected category counts.
A research team calculates expected counts for region and product preference under an independence model.
An AI evaluator examines a contingency table of categorical outcomes, distinguishing observed errors from the counts predicted by a specified comparison model.

Think of it this way

Think of expected frequencies as the seating plan predicted by a probability model. The actual guests may occupy different seats, and the size of that difference needs interpretation against ordinary variation.

Worked example

Consider an invented table of 100 customers. Group A contains 40 people: 30 say yes and 10 say no. Group B contains 60 people: 30 say yes and 30 say no. The column totals are 60 yes and 40 no.
Under independence, the expected counts are A-yes = 40 × 60/100 = 24, A-no = 16, B-yes = 36, and B-no = 24.
The Pearson contributions are 36/24 = 1.5, 36/16 = 2.25, 36/36 = 1, and 36/24 = 1.5. Their sum is 6.25. The ordinary two-by-two independence model has one degree of freedom.
This is the uncorrected Pearson statistic. It is not a causal effect estimate and does not reveal why the groups differ. The observed yes proportions are 75% and 50%; report these descriptive differences alongside any appropriately calibrated test.

Illustrative case study

Fictional case: Harbour Services compares subscription choices across two customer groups. A dashboard initially labels expected frequencies as company targets. The analytics team replaces that wording with “counts under independence” and displays the observed counts separately. It also checks whether customers contributed more than one record before deciding on a test. This invented example shows why both terminology and the observation unit matter.

Common misconceptions and mistakes

Do not replace observed counts with expected counts in a report of what actually happened. Expected values belong to the model.
A fractional expected count is not a data error. A small observed count alone does not determine whether a chi-square approximation is adequate; the expected distribution and design matter.
Agreement with expected counts does not prove a model true. Conversely, a large discrepancy does not by itself establish a practically important association or a causal explanation.

Business use

Expected frequencies help teams understand how categorical results depart from a reference model. They can support quality monitoring, audience analysis, and evaluation of categorical outcomes. Clearly state whether the reference is independence, a contractual distribution, a historical mix, or another hypothesis. This prevents statistical comparisons from being confused with operational goals.

Check your understanding

Why do the worked example's expected counts sum to 100?

Answer: They distribute the fixed total across the model's categories.

What does 24 mean for A-yes?

Answer: It is the fitted count under independence, not an observed total.

Why might repeated customer records matter?

Answer: They can violate the independent-observation framework.

Related terms

Learn next

Learn how marginal totals determine independence-model counts. Then study Pearson residuals and how the sampling design affects the validity of a contingency-table test.

Sources

David Howell, Statistical Methods for Psychology glossary, “Expected frequencies”: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html SciPy, expected_freq documentation: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.contingency.expected_freq.html

Previous: Expected calibration error Next: Expected improvement