Can a statistic be a vector or matrix?

Yes; a vector of means or a covariance matrix is also a function of the sample.

Is a sample statistic always used for inference?

No; many are purely descriptive.

Does changing the sample change the statistic?

Often, although distinct samples can coincidentally produce the same value.

AI Glossary / Statistics and probability / Intermediate

Sample statistics

Definition

Sample statistics are quantities calculated from observed sample data, such as a mean, median, proportion, variance, or correlation. They summarize the sample and may also serve as estimators of population quantities under an appropriate design or model. Formally, a statistic is a function of the observed data and known constants, rather than a formula requiring an unknown population parameter.

What it means

A statistic compresses data into a quantity chosen for a purpose. An average summarizes location, a variance summarizes spread, and a proportion summarizes how often a defined condition occurs. No single statistic describes every important feature of a dataset.
The distinction between statistic and parameter concerns their roles. A sample mean is calculated from observed cases; a population mean describes the specified population. The sample quantity can estimate the population quantity, but the two are not automatically equal.

How it works

Define the sample, variable, and intended summary. Specify the observation unit, units of measurement, relevant dates, and inclusion rules. A statistic based on transactions can answer a different question from one based on customers even if both use the same underlying records.
Choose a calculation suitable for the data and decision. Means respond to extreme values; medians summarize an ordered midpoint; proportions require a clear numerator and denominator. A variance uses squared units, while a standard deviation returns to the original units.
Document weighting and missing-data rules. Weighted and unweighted statistics can differ because they give observations different contributions. Omitting incomplete cases can also change which population the available sample resembles.
Distinguish the rule from its realized value. Before collecting a random sample, a statistic is a random quantity because different possible samples produce different results. After observing the data, its realized value is a specific calculated number.
Use a sampling distribution to study uncertainty when making inference. The statistic's standard error describes variation across repeated sampling under a stated design or model. It is not generally the same as the spread of individual observations.
Assess estimator properties only relative to assumptions and a target. Unbiasedness means the estimator's expectation equals the parameter under the relevant sampling mechanism. Consistency concerns convergence as information grows. Neither property guarantees that one realized estimate equals the truth.
For descriptive work, a statistic can accurately summarize the recorded cases even when broader inference is unsupported. A precise sample proportion from a self-selected poll remains a description of those respondents unless additional assumptions justify population claims.
Keep important context and distributions available. Aggregation can hide subgroups, skewness, extremes, or changes in case mix. In AI evaluation, a mean loss or accuracy statistic should be accompanied by enough detail to understand what cases contributed and where the model performs differently.

Three practical examples

Customer research: an analyst reports a satisfaction proportion for the respondents and explains the survey design before generalizing.
Process analysis: engineers summarize sample dimensions using both location and spread.
AI evaluation: researchers compute test accuracy and subgroup statistics, recognizing that each has sampling uncertainty.

Think of it this way

Think of a statistic as a selected gauge on a dashboard. One gauge may show average speed and another fuel use. Both are useful summaries, but neither records the whole journey. Choosing the gauge and knowing how it was calculated are part of interpreting its reading.

Worked example

Five fictional observations are 2, 4, 4, 6, and 9 units. Their total is 25, so the sample mean is 25/5 = 5. The median is 4, and the range is 9−2 = 7 units.
Deviations from the mean are −3, −1, −1, 1, and 4. Their squares total 9+1+1+1+16 = 28. Using denominator n−1, the sample variance is 28/4 = 7 squared units, and the sample standard deviation is sqrt(7), approximately 2.646 units.
If the event of interest is a value of at least 6, two of five observations qualify, giving a sample proportion of 2/5 = 0.40.
These are different statistics from the same sample. None reveals the population mean, variance, or proportion with certainty. A different selected sample could change every reported value.

Illustrative case study

Fictional case: Willow Retail summarizes an AI-assisted customer service trial with a single average satisfaction score. A closer review shows a wide distribution and different patterns across support channels.
The analyst retains the mean but adds the median, a clearly defined low-score proportion, and subgroup counts. The report distinguishes the observed trial participants from the wider customer population and explains what additional evidence would support broader claims.

Common misconceptions and mistakes

A statistic is not automatically a population truth.
Descriptive accuracy does not establish that the sample represents the intended population.
A standard deviation and a standard error answer different questions.
A numerical summary can be correct while answering the wrong question because the denominator or unit is inappropriate.
An unbiased estimator can still be far from its target in one sample. Conversely, a statistic can be useful descriptively without being an unbiased estimator of a chosen population parameter.

Business use

Use sample statistics to make data understandable, but connect each number to its definition and decision. Require clear denominators, units, weights, and exclusions in recurring business reports. When AI performance claims rely on a sample statistic, present its evaluation scope and uncertainty rather than treating a single displayed score as a permanent system property.

Check your understanding

Which worked-example statistic is expressed in squared units?

Answer: The sample variance.

Why are the mean and median different?

Answer: They summarize location in different ways and respond differently to the values.

Does the observed 40% prove the population proportion is 40%?

Answer: No; generalization requires a design or model and allowance for uncertainty.

Related terms

Sample mean
Sample variance
Sampling distribution
Standard error
Population parameter
estimator
sample proportion

Learn next

Calculate several summaries from one small dataset and explain what each preserves or hides. Then describe how you would assess uncertainty if the data were obtained through a particular sampling design.

Sampling design

Sources

Howell, Statistical Methods for Psychology glossary, Sample statistics and Statistics: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html NIST/SEMATECH e-Handbook, Measures of Location: https://www.itl.nist.gov/div898/handbook/eda/section3/eda351.htm The five-observation calculations and illustrative case are original and fictional.

Previous: Sample Standard Deviation Next: Sample unique