Can a stratified sample be self-weighting?

Yes, for example when the sampling fraction is the same across strata and no later adjustment changes the weights.

Does self-weighting eliminate sampling error?

No; estimates still vary across possible samples.

Can equal household probabilities imply unequal person probabilities?

Yes, if selection within households gives people different overall chances.

AI Glossary / Statistics and probability / Advanced

Self-weighting sample

Definition

A self-weighting sample has equal weights, or raising factors, for the estimator under consideration. In the usual design-weight setting, equal inclusion probabilities give sampled units the same inverse-probability weight. The qualification about the estimator matters: equal initial design weights need not remain equal after other estimation adjustments.

What it means

For an ordinary weighted mean or proportion, a common weight cancels between numerator and denominator. This can simplify calculation. For a population total, however, the common expansion factor still matters; the sample total is not automatically the population total.

Self-weighting describes weighting, not a guarantee of representativeness, independence or negligible error.

How it works

For a probability sample, the design weight for a unit is one divided by its inclusion probability. If every relevant unit has inclusion probability pi, every design weight is 1/pi.

Simple random sampling of a fixed number of units from a finite population provides one example. A stratified design with the same sampling fraction in every stratum can also have equal design weights. Some multistage designs coordinate selection probabilities across stages to achieve the same result.

Inspect the weights actually used for the intended estimate. Adjustments for nonresponse or calibration to population information can change the equality present at selection. Equal weights also leave stratification and clustering relevant to uncertainty calculations.

Three practical examples

1. Customer survey: sample the same fraction of customers from each service category, producing equal initial design weights despite different category sizes.

2. Quality inspection: select a simple random sample of items from a complete production list. Each selected item has the same inclusion probability and expansion factor.

3. Multistage survey: coordinate selection of larger units and sampling within them so that final units have equal overall inclusion probabilities.

Think of it this way

Imagine each sampled record carrying a card saying how many population records it represents in the estimator. In a self-weighting calculation, those cards show the same number. That simplifies adding contributions, but does not say the records were collected independently or that every relevant group responded.

Worked example

Consider a fictional population divided into two nonoverlapping strata. Stratum A contains one hundred customers and stratum B contains three hundred. Select ten customers by simple random sampling without replacement from A and thirty independently from B. Suppose all selected customers respond.

The inclusion probability is 10/100 = 0.1 in A and 30/300 = 0.1 in B. Each design weight is therefore ten. The forty selected customers form a self-weighting sample for the usual inverse-probability total estimator.

Suppose four sampled customers in A and twenty-four in B report using a feature. The estimated population total using it is 10 × 4 + 10 × 24 = 280. The weights sum to 10 × 40 = 400, matching the known population size.

The estimated proportion is 280/400 = 0.7, or 70%. The unweighted sample proportion is also (4 + 24)/40 = 70%. The common factor cancels. However, the unweighted count of twenty-eight users is still a sample count, not the estimated population total of 280.

Now change the design to select twenty customers from each stratum. The inclusion probabilities become 20/100 = 0.2 and 20/300 = 1/15. The design weights are five and fifteen, so this allocation is not self-weighting for the same total estimator.

For this second illustration, suppose eight of A's twenty sampled customers and sixteen of B's twenty use the feature. Their stratum sample proportions are again 40% and 80%.

The weighted total is 5 × 8 + 15 × 16 = 280, and the weighted population count is 5 × 20 + 15 × 20 = 400. The weighted proportion remains 70%. The unweighted proportion is instead 24/40 = 60%, because equal sample counts give the smaller stratum too much influence relative to its population size.

These are hypothetical sample outcomes chosen for transparent arithmetic. They do not imply that repeated samples produce identical estimates or that 70% is the population's true proportion.

In the first design, further adjustments could also break equal weighting. If response differed substantially between strata, simply retaining weight ten for every respondent would not generally address that loss. Any nonresponse adjustment would need explicit assumptions and appropriate information. The self-weighting property at selection does not settle the final estimation problem.

Illustrative case study

Fictional Willow Research receives the first survey's results and sees equal weights. Its analyst correctly uses the unweighted proportion as the numerical equivalent of the design-weighted proportion, but initially applies a simple-random-sample uncertainty formula without examining the stratification.

The reviewer keeps the point estimate and asks for uncertainty calculated under the actual design. Willow documents stratum sizes, sample sizes and response handling. This invented case shows why simplifying a point estimate does not justify simplifying every part of the analysis or assuming the design effect equals one.

Common misconceptions and mistakes

“Self-weighting means weights are one.” They may share any appropriate common value; the worked design uses ten.

“Equal samples per stratum give equal weights.” Equal sampling fractions are the relevant feature in the stratified example.

“Equal weights make the design a simple random sample.” Other designs can have equal weights while different joint selection probabilities affect variance.

“No weighting work is needed later.” Nonresponse and calibration can require further adjustments.

Business use

Ask which population, sampling unit and estimator the self-weighting claim concerns. Verify inclusion probabilities across all stages and strata. Distinguish original design weights from the final analysis weights supplied with a dataset.

For AI-generated survey summaries, check whether totals require expansion even when proportions can be computed without unequal weights. Retain the information needed for design-aware uncertainty estimates. When comparing surveys, do not assume that two self-weighting samples have the same precision, coverage or response quality.

Check your understanding

Why is the first weight ten?

Answer: It is the reciprocal of the inclusion probability 0.1.

Why do weights cancel for the first proportion?

Answer: The same factor multiplies both numerator and denominator.

Why is the second unweighted proportion different?

Answer: It gives equally sized samples from unequally sized population strata equal influence.

Related terms

Stratified sampling
Probability proportional to size
Calibration
Design weight
inclusion probability
nonresponse adjustment
survey variance
raising factor

Learn next

Change the first design's sample sizes to five and fifteen. Recalculate the inclusion probabilities and common weight, then explain why the expected precision may change despite retained self-weighting.

Sources

OECD Glossary of Statistical Terms, Self-weighting sample, estimator-specific definition: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf

Statistics Canada, Statistics: Power from Data!, 3.5.1 Weighting: https://www150.statcan.gc.ca/n1/edu/power-pouvoir/ch6/5214809-eng.htm

Previous: Self-training Next: SELU Function