Stratified sampling in audit: formula and worked example
Stratified sampling splits the population into groups of similar value and draws a random sample from each; the largest items are examined 100%. The spread within each stratum falls, and so does the sample: in this guide's example, from 1,153 invoices to 336.
Updated 10 November 2026
Why stratify
In classical variables sampling the sample size depends on how spread out the values are. A receivables ledger with invoices of 200 and of 300,000 has a huge standard deviation, and a simple random sample would need to be nearly as large as the population to give useful precision. Split the population into strata of similar value and the spread inside each stratum is much smaller; the strata samples add up to a fraction of the unstratified sample.
ISA 530 Appendix 1 sets this out: stratifying reduces the variability within each stratum, so the sample can be smaller without increasing sampling risk, and in tests of details it directs effort to the large items, where the largest overstatement may sit. Two rules from the same appendix matter at evaluation: results from a stratum are projected only to that stratum, and the projected misstatement of each stratum is added up to conclude on the total. AU-C 530 takes the same approach.
You can also stratify by another risk characteristic, such as the age of balances when testing the bad debt allowance. This guide covers stratification by value, the usual choice in variables sampling.
Defining the strata
First set aside the top stratum: items at or above a threshold are all examined. They carry no sampling error, because they are not sampled, and they remove from the population the items that add most of the spread. The threshold is a judgement; a reasonable starting point is the amount at which a single item approaches a meaningful share of tolerable misstatement.
The rest is split with one of these rules:
- Cumulative square root of frequency (Dalenius-Hodges). Put the values into many classes of equal width, count the items in each, cumulate the square root of those counts and cut the total into equal parts. It gives boundaries close to optimal for Neyman allocation and is AuditSampler's default.
- Equal value. Each stratum holds roughly the same total book value. Easy to explain and review, though it usually needs a somewhat larger sample.
- Equal count. The same number of items in each stratum. Simple, but it tends to leave the last stratum with a wide spread.
Three to six strata are usually enough. More strata cut the sample very little and complicate selection and documentation.
Audit variable sampling: the sample size formula
With tolerable misstatement TM and expected misstatement EM, the desired precision (allowance for sampling risk) is A = TM − EM. With z the two-sided normal factor (1.96 at 95% confidence), Nh the number of items and Sh the standard deviation of each sampled stratum:
Neyman: n = (Σ Nh·Sh)² ÷ ((A/z)² + Σ Nh·Sh²)
Proportional: n = N × Σ Nh·Sh² ÷ ((A/z)² + Σ Nh·Sh²)
Neyman allocation spreads the sample in proportion to Nh·Sh: more items where there are more items and more spread. Proportional allocation spreads it by Nh alone, so the strata of small invoices, which are numerous, take most of the sample. Neyman gives the smallest sample for a given precision and is the usual choice in audit.
Everything is rounded up. If a stratum is allocated more items than it has, it is examined in full and the rest is sized again without it. Then a minimum per stratum applies (5 items by default). For planning, AuditSampler uses the standard deviation of each stratum's book values, the prudent mean-per-unit planning figure.
Stratified sampling audit example
Example population: 1,200 open receivable invoices at year end with a book value of 16,967,214.41. They range from 227.28 to 385,903.86, with a mean of 14,139.35 and a standard deviation of 23,687.40. Tolerable misstatement 375,000, expected misstatement 50,000, 95% confidence.
1. Strata
Top stratum: invoices of 150,000 or more. The rest in four strata by cumulative square root:
| Stratum | Values | Items | Book value | Std dev | Sample |
|---|---|---|---|---|---|
| 1 | 227.28 to 6,774.09 | 555 | 1,865,363.65 | 1,724.50 | 75 |
| 2 | 6,774.09 to 15,773.90 | 337 | 3,481,596.70 | 2,543.47 | 68 |
| 3 | 15,773.90 to 35,494.12 | 202 | 4,600,097.03 | 5,448.87 | 87 |
| 4 | 35,494.12 to 97,481.88 | 99 | 5,326,303.04 | 17,237.63 | 99 |
| Top | 150,000 or more | 7 | 1,693,853.99 | — | 7 |
| Total | 1,200 | 16,967,214.41 | 336 |
Each stratum includes its lower bound and excludes its upper one, except stratum 4, which runs up to the highest invoice below the threshold. No invoice falls between 97,481.88 and 150,000.
2. Size with Neyman allocation
- Desired precision: A = 375,000 − 50,000 = 325,000. A/z = 325,000 ÷ 1.96 = 165,819.37; (A/z)² = 27,496,064,628.10.
- Strata 1 to 4: Σ Nh·Sh = 4,621,442.91 and Σ Nh·Sh² = 39,244,523,816.90.
- n = 4,621,442.91² ÷ (27,496,064,628.10 + 39,244,523,816.90) = 320.01, rounded up to 321.
- Stratum 4's share is 321 × 36.93% = 118.5, so 119 items, but it only has 99: it is examined in full.
- Size again on strata 1 to 3: Σ Nh·Sh = 2,914,917.30, Σ Nh·Sh² = 9,828,062,556.99 and n = 227.65, rounded up to 228. Allocated by Nh·Sh: 75, 68 and 87 items.
- Total: 75 + 68 + 87 + 99 + 7 = 336 items. Planned precision is 322,741.48, inside the 325,000 sought.
3. How the design changes the sample
| Design (same TM, EM and confidence) | Sample |
|---|---|
| Unstratified (mean-per-unit) | 1,153 |
| One stratum plus the top stratum | 1,093 |
| Cumulative square root, proportional allocation | 711 |
| Equal value, proportional allocation | 621 |
| Equal value, Neyman allocation | 416 |
| Cumulative square root, Neyman allocation | 336 |
Setting aside only the 7 large invoices barely helps (1,093). What cuts the sample is stratifying the rest and allocating with Neyman.
Evaluating the sample
For each item tested, misstatement is d = book value − audited value (positive for an overstatement). In each sampled stratum it is projected with one of three estimators:
- Difference: Nh × mean difference in the stratum sample.
- Ratio: stratum book value × Σ differences ÷ Σ book value of the stratum sample.
- Mean-per-unit: stratum book value − Nh × mean audited value of the sample.
The variance of each stratum is Nh² × sh² ÷ nh × (1 − nh/Nh), sh being the sample standard deviation (of the differences, of the ratio residuals or of the audited values). Strata examined 100% count at their actual misstatement with no variance. Precision is z × √(Σ variances), and the population is accepted when |projected misstatement| + precision does not exceed tolerable misstatement.
Results of the example
The 336 invoices are selected at random (seed 2026) and tested. 8 invoices show a difference, with known misstatement of 133,430.37: 4 in stratum 3 (52,967.78), 3 in stratum 4 (61,352.95) and 1 in the top stratum (19,109.64). Strata 1 and 2 show none.
| Estimator | Projected misstatement | Precision (95%) | Upper limit | Conclusion |
|---|---|---|---|---|
| Difference | 203,445.25 | 118,654.39 | 322,099.64 | Accept |
| Ratio | 201,532.87 | 117,843.53 | 319,376.40 | Accept |
| Mean-per-unit | 188,307.14 | 330,098.26 | 518,405.40 | Reject |
With difference estimation, stratum 3 projects 202 × 52,967.78 ÷ 87 = 122,982.66, with a standard error of 60,539.06. Stratum 4 and the top stratum add their actual 80,462.59 with no sampling error. Total: 203,445.25, and precision is 1.96 × 60,539.06 = 118,654.39. The 322,099.64 limit is below 375,000. Ratio estimation gives almost the same answer because the mean book value of the stratum 3 sample (23,132.47) is very close to that of the whole stratum (22,772.76).
Mean-per-unit rejects on the same sample. It does not find more errors: its precision depends on the spread of audited values, not of the differences, and comes out nearly three times wider. That is why, when the client has a reliable listing of book values, you use difference or ratio estimation and keep mean-per-unit for when there is no book value per item.
Before concluding. Projected misstatement (203,445.25) is well above the 50,000 expected at planning. ISA 530 paragraph A22 warns that in that case the risk that actual misstatement exceeds tolerable misstatement may be unacceptable even though the limit is inside. The 8 differences are investigated, the known ones proposed as adjustments and the projected amount carried to the summary of misstatements.
Common mistakes
- Projecting everything with a single rate instead of stratum by stratum.
- Sampling the top stratum instead of examining it in full, or not reconciling the strata total to the ledger.
- Trusting the precision with very few differences. With difference or ratio estimation, when there are hardly any differences the normal approximation is unreliable; here a single stratum with 4 differences provides all the variance.
- Using mean-per-unit when book values exist: much worse precision for the same sample.
- Redefining strata after selection, or selecting items that fall in no stratum because the population has changed.
- Ignoring the minimum per stratum and leaving a stratum with one or two items and no standard deviation to compute.
Stratified sampling in AuditSampler
AuditSampler Pro defines the strata (cumulative square root, equal value or equal count) with an optional top stratum, sizes the sample with Neyman or proportional allocation, selects at random within each stratum with a reproducible seed and evaluates with difference, ratio or mean-per-unit estimation. Results export to CSV; a PDF working paper is not yet available for this method.
Frequently asked questions
What is stratification in audit sampling?
Dividing the population into sub-populations of similar items, usually by monetary value, and sampling each one. It reduces the spread within each stratum and so the sample size needed for the same sampling risk.
How do you calculate the sample size in stratified sampling?
With Neyman allocation, n = (Σ Nh·Sh)² ÷ ((A/z)² + Σ Nh·Sh²), where A is tolerable minus expected misstatement and z the normal factor (1.96 at 95%). The sample is split between strata in proportion to Nh·Sh.
What is the difference between Neyman and proportional allocation?
Proportional allocation splits the sample by the number of items in each stratum; Neyman by the number of items times the standard deviation. Neyman sends the sample to the strata with more spread and needs fewer items: 336 against 711 in the example.
What is a top stratum?
The group of items above a threshold that are all examined. It has no sampling error and its misstatement is added as it is to the projected misstatement of the other strata.
Which estimator should I use to evaluate a variables sample?
Difference estimation when misstatements do not depend on item size, ratio estimation when they grow with it, and mean-per-unit only when there is no reliable book value per item. Difference and ratio need enough differences in the sample.
- IAASB, ISA 530 Audit Sampling (Appendix 1, paragraphs 1 to 4; A22)
- AICPA, AU-C Section 530 Audit Sampling
- AICPA, Audit Guide: Audit Sampling (classical variables sampling)
- Cochran, W. G., Sampling Techniques, 3rd ed., Wiley (Neyman allocation, cumulative square root stratification)
Strata, sample and evaluation in one place
AuditSampler Pro defines the strata, sizes the sample with Neyman allocation, selects it with a reproducible seed and evaluates with difference, ratio or mean-per-unit estimation. Client data stays on your computer.