Abstract
The severity of punishment is the “neglected sibling” of the Allingham–Sandmo deterrence model: credible field-experimental evidence on whether taxpayers respond more to being caught or to being punished has been scarce because governments rarely randomize sanctions. We address this gap with a large-scale randomized experiment run by the Norwegian Tax Administration, studying taxpayers who reduce pre-filled wage income during electronic filing. We randomize three enforcement instruments: a real-time digital prompt when the taxpayer edits wages, an audit when a reduction is filed, and the timing of penalty-tax assignment among those who qualify. For the penalty arm, some receive the penalty immediately; others experience an administrative delay that is not announced ex ante. This design identifies the causal effect of realized punishment on subsequent reporting, conditional on detection. The effect of punishment is large: treated taxpayers pay roughly 12,000 NOK more in income tax and report roughly 17,000 NOK more in wages the next year, relative to those whose penalty was delayed. Audits alone have only small effects on next-year reporting. The digital prompt cuts the share of filers who submit a wage reduction by roughly half. Effects of the penalty are much larger among previously noncompliant taxpayers and are concentrated on wage reporting rather than on deductions. The marginal deterrent on this margin is the sanction rather than detection.