How large is the downward bias in household survey income estimates?
Core argument: Total survey error in the Current Population Survey from 1984 to 2022 is almost always negative for both recipients and dollars, indicating survey means are biased downward, with bias of 40% or more for nearly half the variables.
To directly measure bias in survey estimates, we calculate the difference between the weighted survey estimate and the survey target constructed from published totals [TSE or total survey error], using public-data adjustments for intentional coverage differences for average dollars received and the recipient share of the population across our income sources. Figures 1 and 2 summarize TSE in the CPS from 1984 to 2022, expressed as a share of the survey target for our measures of recipients (seven income sources) and dollars (eight income sources). In levels, TSE is almost always negative for both recipients and dollars (with SSI dollars an exception in recent years), indicating that survey means are systematically biased downward. For nearly half of the variables the bias is 40% or more. Yet, there is substantial heterogeneity across income sources. TSE tends to be more modest for SSA programs (specifically OASDI and OASI, for which the net bias is below 15% for recipients and dollars), while it is much larger for income sources such as pensions, UI, and SNAP (for which the net understatement exceeds 40%).



