Panel Conditioning Biases in the Current Population Survey's Food Security Supplement

We estimate the extent to which the methodological problem called panel conditioning biases the federal government's estimates of the prevalence of food insecurity in the United States. To do so, we use 2002 through 2020 data from the Current Population Survey's Food Security Supplement-the same data used to produce the federal government's annual statistics about food insecurity. We take advantage of the CPS's rotating panel design feature to estimate the effects of panel conditioning. By comparing CPS respondents who participated in the Food Security Supplement in each of two consecutive years but who-strictly by chance-were selected to begin the CPS one year apart, we both approximate an experimental design and derive estimates of panel conditioning that are purged of biases from panel attrition. For the 200,000+ unique households in our sample, the treatment is having previously participated in the Food Security Supplement; the outcome is participants' subsequent responses to survey questions about food security. We find that in nearly every year people in the treatment group-that is, the group of people who have previously responded to the Food Security Supplement-are less likely to be food insecure than people responding for the first time. These differences are statistically significant and large in magnitude. We conclude that the federal government's estimates of the prevalence of food insecurity in America are substantially biased; depending on the mechanism underlying panel conditioning, the true prevalence of food insecurity may be substantially higher or lower than officially reported.