Benford's Law in Public Sector Audit: Uses and Limits

Benford's law describes the expected frequency of leading digits in many naturally occurring sets of numbers: roughly thirty per cent begin with 1, declining to under five per cent beginning with 9. Data sets spanning several orders of magnitude and arising from multiplicative processes tend to conform. Because fabricated figures often do not, digit analysis became a standard item in the forensic accounting toolkit — and, subsequently, a standard source of misinterpretation.

Where it applies

Conformity requires specific conditions: the data should span multiple orders of magnitude, arise without an imposed minimum or maximum, and not be assigned or sequential. In public finance the data sets that plausibly qualify include disbursements across a full year, journal entry amounts, purchase card transactions, grant expenditure detail, and utility billing amounts.

Where it does not, and why the exceptions matter more

The following will fail a conformity test for entirely innocent reasons, and each failure will consume investigative time if the analyst has not anticipated it:

• Amounts subject to a threshold. Any population filtered at, say, transactions over one thousand dollars has an imposed minimum and will not conform.
• Assigned numbers. Cheque numbers, invoice numbers, account codes, parcel identifiers.
• Contractual or scheduled amounts. Debt service, lease payments, insurance premiums, salary lines — a small set of values repeated many times.
• Narrow-range populations. Per diem reimbursements, permit fees, and any fee schedule.
• Populations dominated by one payee. A large recurring payment to a single vendor will distort the distribution on its own.

Roughly speaking, most of a public entity's general ledger falls into one of these categories. The subset to which the test legitimately applies is smaller than the enthusiasm for it suggests.

Using it correctly

The productive application is not "does this population conform" but "which subsets deviate most". Partition the disbursement population by department, by vendor, by approver, and by period, then rank subsets by mean absolute deviation. The ranking becomes an input to audit sample selection. A department whose disbursements deviate substantially from the pattern of comparable departments is worth a closer look — not because deviation indicates fraud, but because it indicates something the analyst does not yet understand.

Second-order tests are often more informative than the first-digit test. The last-two-digit test detects rounding and fabrication tendencies; a spike at 00 usually reflects estimates or round-number contracting rather than misconduct, but it identifies where estimation is occurring. Duplicate-value and "number frequency" tests — which values appear far more often than expected — frequently produce the most actionable output of any digit analysis, and require no distributional assumptions at all.

What it can never establish

Digit analysis identifies populations for examination. It does not identify transactions, it does not identify individuals, and it does not constitute evidence of impropriety. A finding phrased as "the department's disbursements do not conform to Benford's law" is, standing alone, unsupportable as a conclusion and will be dismantled on review. The defensible phrasing is that the analysis informed sample selection, followed by whatever the substantive testing found.

Practical implementation

The computation is a dozen lines of code or a handful of spreadsheet formulas. The work is entirely in data preparation: excluding reversals and their originals, handling credits, converting to absolute values, removing the categories listed above, and documenting each exclusion. An undocumented exclusion is the fastest way to make an analysis unreproducible, and reproducibility is the only property that makes the exercise worth performing at all.