Overview
Benford's law says that in many real data sets, about 30.1% of numbers start with the digit 1, and far fewer start with 9. Auditors use it to spot figures that may have been invented.
Enter the share of your numbers that start with 1, and how many numbers you checked.
Worked examples
| Inputs | Result |
|---|---|
| Observed Leading Digit 1 Share (%)27.8Sample Size1,200 | Observed leading digit 1 share: 27.8%Benford expected: 30.1%z-score: -1.74p-value: 0.08 |
Show 2 more examples
| Observed Leading Digit 1 Share (%)30Sample Size500 | Observed leading digit 1 share: 30%Benford expected: 30.1%z-score: -0.05p-value: 0.96 |
| Observed Leading Digit 1 Share (%)22Sample Size2,000 | Observed leading digit 1 share: 22%Benford expected: 30.1%z-score: -7.9p-value: 0 |
Observed leading digit 1 share by observed leading digit 1 share
With the calculator's starting values, the observed leading digit 1 share is 27.8%. The table shows how it changes with the observed leading digit 1 share.
| Observed Leading Digit 1 Share (%) | Observed leading digit 1 share |
|---|---|
| 15% | 15% |
| 20% | 20% |
| 25% | 25% |
| 27.8%starting value | 27.8% |
| 30% | 30% |
| 35% | 35% |
| 40% | 40% |
Other inputs: Sample Size: 1,200. Change them in the calculator above for your own numbers.
How it works
The expected share for a leading 1 is log₁₀(2), about 30.1%. The z-score measures how far your share is from that, given your sample size, and the p-value is the chance of a gap that large by luck. This checks the digit 1 only. A full test compares all nine digits with a chi-square test.
The formula
| n | Number of values checked |
Practical tips
- 27.8% across 1,200 numbers gives z ≈ −1.74 and p ≈ 0.08, which isn't significant at 5%.
- Benford's law suits data spanning several orders of magnitude, such as invoices, populations or stock prices.
- Assigned numbers, such as phone numbers or prices capped in a narrow range, don't follow it.
Frequently asked questions
Does a mismatch prove fraud?
No. It's a signal to look closer, since many honest data sets don't fit.
What are the expected shares for other digits?
log₁₀(1 + 1/d): 17.6% for 2, 12.5% for 3, down to 4.6% for 9.
How many numbers do I need?
At least a few hundred for a meaningful test.