Does the method require knowing every population value?

No. It uses sampled values and population counts for the relevant strata.

Are all survey weights the same?

Only within the simple stratum-specific construction described here; weights can differ across strata or in more complex designs.

Does increasing a respondent's weight reduce uncertainty?

No. A larger expansion factor does not add observations or guarantee greater precision.

AI Glossary / Statistics and probability / Intermediate

Number raised estimation

Definition

Number raised estimation is a survey estimation method that expands sample observations using a common population-to-sample count ratio within each sampling stratum. In its basic form, a sampled unit receives weight N/n, where N is the number of population units in that stratum and n is the number observed. Multiplying observations by these weights and adding them estimates a population total. The simple formula belongs to an appropriate equal-probability sampling framework.

What it means

The weight describes how many population units an observed unit represents for estimation purposes. It does not mean that those unobserved units have been individually measured or are identical to the observed unit.

Different strata can have different weights because their sampling fractions differ. A small stratum sampled intensively may give each observation less weight than a large stratum sampled lightly. This matters when combining survey results: the average of all responding records can differ substantially from the estimated population average.

How it works

Define the target population, sampling units, and any strata. Obtain reliable population counts for each stratum and draw samples using the intended probability design. In a complete-response stratified simple random sample, calculate each stratum's expansion factor as its population count divided by its selected sample count.

Multiply each observed value by the relevant factor and sum across strata. For a population mean, divide the estimated total by the known total population count. Calculate sampling uncertainty using formulas or survey software that account for the design and sampling fractions.

The historical glossary definition uses responding sample counts. Substituting respondent counts for selected counts requires care: it does not automatically eliminate nonresponse bias. Representativeness within adjustment groups needs justification. Unequal selection probabilities, clustering, frame errors, and missing outcomes may require a more elaborate estimator and variance calculation.

Three practical examples

An organization estimates total training hours across branches by expanding a random sample within each branch-size group.
A business survey estimates total equipment spending using known counts of firms and equal-probability samples within industry strata.
An AI adoption survey combines small and large organizations using their population representation rather than simply averaging all questionnaire records.

Think of it this way

Think of sampling spoonfuls from several well-mixed containers. Each spoonful represents part of its own container, so its contribution depends on that container's size and how much was sampled. Taking more spoonfuls from a small container should not make that container dominate the estimate for the entire collection.

Worked example

Suppose a fictional population contains 40 small offices and 10 large offices. A simple random sample selects four small offices with weekly training hours of 10, 14, 18, and 22, and two large offices with 30 and 34 hours. Assume complete response and accurate counts.

The small-office weight is 40/4 = 10. Their observed total is 64, producing an estimated population total of 640 hours. The large-office weight is 10/2 = 5. Their observed total is also 64, producing an estimated total of 320 hours.

The combined estimate is 960 hours across 50 offices, or 19.2 hours per office. The unweighted sample mean is 128/6, approximately 21.33 hours. It overrepresents the large-office stratum relative to its population share. The difference illustrates weighting, not proof that the estimate equals the unknown population truth.

Illustrative case study

Fictional case study: Maple Workforce surveys organizations about staff time spent learning AI tools. Its analysts deliberately sample a larger fraction of large organizations to understand their varied training practices. An initial dashboard averages all responses equally and presents that number as the population mean.

The survey lead rebuilds the estimate using population counts and the appropriate stratum weights. The team also reviews nonresponse patterns, documents frame coverage, and adds uncertainty intervals calculated from the sampling design. It labels its result an estimate and keeps weighted totals separate from actual response counts. This fictional example illustrates responsible aggregation; it does not report a real survey outcome.

Common misconceptions and mistakes

The most common mistake is assuming that expansion creates new information. A weighted estimate representing thousands of units may still be based on a modest number of observations.

Another mistake is interpreting the method's design-unbiasedness as a promise that every individual estimate is correct. The property concerns repeated sampling under specified conditions. Number raising does not repair systematic nonresponse, inaccurate measurements, or excluded population groups by itself. Finally, equal weights within a stratum should not be applied blindly to a design with unequal inclusion probabilities.

Business use

Number raised estimation helps businesses turn carefully designed survey samples into estimates of market totals, organizational practices, resource use, and adoption rates. It is particularly transparent when reliable population counts exist but useful auxiliary totals for more elaborate estimation are unavailable.

For AI-assisted reporting, store the sampling design, population counts, response counts, and weighting rules alongside the results. Generated summaries should distinguish the number surveyed from the population represented and should not present weighted counts as additional respondents.

Check your understanding

A stratum contains 300 units and a complete random sample of 30 units has an observed total of 450. What is the number-raised total estimate?

Answer: The weight is 300/30 = 10, so the estimated total is 4,500. This remains an estimate subject to the sampling design and data quality.

Related terms

Stratified sampling
Sampling error
Ratio estimation
Horvitz - Thompson estimator
Survey weight
expansion estimator
simple random sample
population total
nonresponse adjustment

Learn next

Study probability sampling and inclusion probabilities, then compare expansion estimators with ratio and calibration methods. Practice calculating design-based uncertainty and explaining why weighting cannot substitute for good coverage, response, and measurement.

Sources

OECD Glossary of Statistical Terms, entry for Number raised estimation, gives the historical survey definition: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf

Australian Bureau of Statistics, Basic Survey Design: Analysis, discusses number-raised estimation: https://www.abs.gov.au/websitedbs/d3310114.nsf/home/Basic%20Survey%20Design%20-%20Analysis

The example is original. Unbiasedness is conditional on the sampling and response assumptions described above.

Previous: Number needed to treat Next: Numeric prediction