Research

Abstract

Repeated interviewing is useful for measuring belief dynamics, but it may also change the answers that household surveys elicit. This paper compares panel conditioning across three major expectation surveys: the ECB Consumer Expectations Survey (CES), the Bundesbank Online Panel–Households (BOP-HH), and the New York Fed Survey of Consumer Expectations (SCE). I harmonize point forecasts, subjective density forecasts, uncertainty, internal consistency, and ex-post accuracy and estimate exact-tenure profiles over the first twelve interviews. Repeated participation systematically reduces subjective uncertainty and point–density inconsistency, but the magnitude and shape of these changes are not survey-invariant. The SCE exhibits the largest declines in point expectations, inconsistency, and forecast errors; BOP-HH is often closer to CES Germany in belief levels and forecast errors, but closer to the SCE in uncertainty. Predicted initial conditions explain an important share of several cross-survey gaps, while substantial tenure dynamics remain after the first interviews. The BOP-HH participation design provides additional mechanism evidence: scheduled breaks partially reverse the decline in uncertainty, whereas randomly assigned prior exposure to the same item has little effect conditional on general survey tenure. A randomized wording experiment produces large contemporaneous shifts in reported beliefs and reference objects, but these effects largely disappear in subsequent interviews. Taken together, the results indicate that measured panel conditioning reflects a combination of initial conditions, survey familiarity, and elicitation design rather than a single survey-invariant learning process.

Abstract

Fornaro and Wolf (2026) show that AI-driven automation can redistribute income from labour to capital and, because workers have a higher marginal propensity to consume, compress aggregate demand: an “AI slump.” This paper asks whether UK occupations with greater ex ante exposure to generative AI experienced differential changes in realised wages and employment after ChatGPT’s November 2022 release. Using official UK earnings and employment data (2016–2025), continuous and categorical exposure designs, and a diagnostic sequence combining conventional and synthetic difference-in-differences, doubly robust DiD, joint pre-trend and placebo tests, and Honest-DiD sensitivity analysis, we find no stable exposure-graded decline in wages or employed-person headcounts; a positive employee-jobs signal survives in one closely related specification family, but its post-2022 timing remains uncertain. These findings concern an occupational labour-market margin and do not directly identify aggregate demand, consumption, or labour-share redistribution.

Abstract

How does public health insurance affect households’ ability to save out of temporary income gains? This paper studies the effect of Medicare eligibility on households’ marginal responses to income shocks using microdata from the Survey of Consumer Expectations. Exploiting the sharp eligibility threshold at age 65 in a regression discontinuity design, the analysis shows that Medicare eligibility increases the saving response to positive income shocks by approximately 8 percentage points. By contrast, responses to negative income shocks change little at eligibility. Auxiliary evidence from the Health and Retirement Study indicates that Medicare substantially reduces exposure to large out-of-pocket medical expenditures, particularly in the upper tail of the expenditure distribution. The findings suggest that public health insurance affects household financial behavior not only through direct expenditure reductions, but also by changing households’ ability to carry forward transitory income gains through saving.