Is it the same as a primary sampling unit?

Only in a single-stage design or another situation where those roles coincide.

Does a row in an analysis file identify the final sampling unit reliably?

Not without design information.

AI Glossary / Statistics and probability / Intermediate

Ultimate sampling unit

Definition

An ultimate sampling unit is a unit selected at the final stage of a sampling design. In a multistage household survey, it might be a household or an individual, depending on whether selection stops at households or continues to people within them. It is the last unit subject to sample selection, not automatically the smallest object observed or the unit used in every analysis. The term is meaningful only in relation to the actual sampling design.

What it means

Large surveys often select areas first, addresses or households next, and sometimes individuals last. Each stage narrows the sample. The ultimate unit identifies where that selection process ends. A selected household can supply information about all its members without every member having been separately sampled at an additional stage. Conversely, if one member is randomly selected within each chosen household, that person is the ultimate sampling unit. Keeping this distinction clear prevents confusion between selection, reporting and analysis.

How it works

Describe the sampling stages in order, identify the units eligible at each stage and record their conditional selection probabilities. For a unit reached through a unique selection path, its inclusion probability is the product of the relevant conditional probabilities along that path. A basic design weight is the reciprocal of its inclusion probability. More complicated designs, including multiple paths to selection, need the corresponding inclusion-probability calculation. Analysis must also reflect clustering and stratification; simply knowing the final unit does not determine a valid variance estimator.

Three practical examples

Example 1: A survey selects districts and then households, interviewing every adult in each selected household; households are the ultimate selected units. Example 2: A similar survey randomly selects one adult within each household; adults are now the ultimate units. Example 3: An AI adoption study selects companies, then branches, then employees; employees are the final units when that last selection stage is used. The questionnaire respondent and the ultimate unit may coincide, but they need not.

Think of it this way

Think of a sequence of draws from progressively smaller containers. The last draw determines the ultimate sampling unit. Looking at several characteristics inside the selected item does not necessarily add another sampling stage. What matters is whether another random selection actually occurs, rather than how detailed the resulting measurements become.

Worked example

Suppose a population has ten districts, each containing 100 households, and every household contains two eligible adults. Select two districts by simple random sampling, then ten households within each selected district, then one of the two adults within each selected household. For a particular adult, the inclusion probability is (2/10) times (10/100) times (1/2), equal to 0.01. The basic weight is therefore 100. There are 20 selected adults, while the population contains 2,000 adults, and the weights sum to 2,000. The adult is the ultimate sampling unit. If the last stage is replaced by observing both adults in each selected household, households become the ultimate selected units, and each adult's inclusion probability becomes 0.02. The example assumes equal household sizes, full coverage and response, and the stated selection procedures.

Illustrative case study

In this fictional case study, Birch Research surveys employee use of AI tools. Its team initially labels branches as the final sampling units, but the fieldwork plan also randomly chooses employees within branches. Correcting the design description clarifies that employees are the ultimate units and that employee selection probabilities must enter the weights. The team retains branch identifiers so analysts can account for clustering. This documentation prevents a later analyst from treating the employee records as a simple random sample.

Common misconceptions and mistakes

Common mistakes include equating the ultimate unit with the first-stage cluster, assuming it is always an individual, and identifying it from spreadsheet rows alone. Another mistake is dropping earlier-stage probabilities when calculating weights. The final-stage selection fraction is not generally the overall inclusion probability. Nonresponse and coverage problems also remain relevant after the sampling stages are correctly described. Finally, treating all final records as independent can understate uncertainty when units share sampled clusters.

Business use

Businesses commissioning surveys should request a clear sampling diagram and definitions of the population, selection units, observation units and reporting units. Preserve identifiers and probabilities needed for weighting and variance estimation. If a vendor changes the final stage, ask how that affects representation, respondent burden and analysis. The ultimate-unit concept is useful for reviewing design documentation and interpreting what a survey can represent. It does not, by itself, establish that the sampling frame is complete or the responses are unbiased.

Check your understanding

What is the ultimate unit when one adult is selected within each sampled household?

Answer: The adult.

What is the adult's inclusion probability in the worked example?

Answer: 0.01.

What is the corresponding basic weight?

Answer: 100.

Why keep district and household information?

Answer: Earlier stages affect inclusion probabilities and the dependence structure needed for appropriate analysis.

Related terms

Multistage Sampling
Related terms include primary sampling unit
secondary sampling unit
inclusion probability
design weight
observation unit and cluster sampling
but those roles should be documented explicitly

Learn next

Learn next how unequal selection probabilities and nonresponse adjustments change survey weights. Then examine how stratification and clustering enter variance estimation. Reconstruct a small survey's selection path from its documentation and check whether the analysis file preserves the information needed to represent that design correctly.

Variance estimation

Sources

Sources: OECD, Glossary of Statistical Terms, entry Ultimate sampling unit, https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf; Eurostat, Household Budget Survey 2015 EU Quality Report, https://ec.europa.eu/eurostat/documents/54431/1966394/HBS_EU_QualityReport_2015.pdf/72d7e310-c415-7806-93cc-e3bc7a49b596. These primary institutional sources identify final-stage units and survey design practice. The equal-size example is original and independently checked; the company survey case is fictional.

Previous: Łukasiewicz logic Next: Unary code