The American Community Survey (ACS) is the largest recurring representative household survey in the U.S. and an important source of information on household economic well-being. The ACS elicits information on income using a rolling past-12-months reference period, which represents the year-long period ending on the date of survey. However, it is unclear whether this reference period facilitates more accurate responses than alternative reference periods, such as the past-calendar year. To test the effect of the reference period on the accuracy of survey responses, we use a randomized controlled trial in which ACS respondents were assigned to questions that asked about their income using the traditional past-12-months reference period or the prior calendar year as the reference period. Exploiting this random assignment, we identify the effect of the reference period on the accuracy of income data by comparing the agreement of survey responses with administrative records from the Social Security Administration’s (SSA) Payment History Update System (PHUS) and the Census Bureau’s Longitudinal Employer-Household Dynamics (LEHD) program. We find that the calendar-year reference period increases survey-administrative record agreement at the intensive margin but has no statistically significant effect on agreement rates at the extensive margin. We find stronger evidence of a positive and significant effect on intensive margin agreement for earnings than for Social Security. These results suggest that aligning survey reference periods with administrative records, which are typically at the calendar-year period, may improve the quality of income data.