When Experts Expect Information to Travel
The IZA experiment did not only challenge an assumption about respondents; it also challenged a prior held by people who study retirement, aging, and longevity. This analysis compares what 262 academic experts predicted with what the information treatment actually produced, then places the result beside labor-market and gender statistics that change the stakes without changing the causal evidence. The practical judgment is twofold: expert forecasts can be accurate about averages while missing a crucial behavioral bridge, and a failed bridge can matter differently for people with different work and pension exposures.
The comparison is useful because the experts were not asked to guess randomly. They saw profiles based on real survey observations and forecast knowledge, own-longevity beliefs, and the effect of receiving actuarial information. Their predictions make an implicit theory visible: if people’s estimates of similar others are corrected, their own estimates should move in the same direction. The data show where that theory worked and where it did not.
What experts predicted—and missed
The expert survey was conducted before the household results were known. The final sample contained 262 academic experts, recruited from researchers working in retirement, aging, longevity, behavioral economics, demography, finance, and related areas. Each expert assessed constructed profiles based on the individual survey. The exercise was not another household experiment. It was a way to measure collective expectations about how respondents would answer and how they would react to the information treatment.
On the descriptive questions, the experts performed reasonably well. They anticipated that respondents would be somewhat pessimistic about the longevity of similar people, that pension and financial knowledge would be limited, and that the knowledge domains would not line up neatly. They also predicted average own-longevity beliefs with striking accuracy: about 85.3 years for untreated profiles, essentially the same as the observed mean.
Their forecasts of the knowledge structure were close in several specific ways. Experts predicted a population-longevity bias of about −1.84 years, compared with an observed bias of roughly −2.41 years. They predicted an average correct-answer rate of 0.39 in the pension module, above the observed 0.26, and 0.45 in the financial module, close to the observed 0.41. For the relationships among the domains, they predicted a longevity–financial correlation of −0.05, matching the data; a longevity–pension correlation of 0.06, slightly above the observed 0.00; and a pension–financial correlation of 0.20, somewhat below the observed 0.29.
Those results matter because they establish that the experts understood much of the landscape. They did not imagine a uniformly informed population. They captured the broad fragmentation of knowledge and the personal optimism in own-longevity expectations. The later failure to predict updating therefore cannot be dismissed as a general inability to understand the respondents.
The major miss appeared when experts were asked to forecast the treatment effect. For profiles with an average population-longevity bias of about −2.4 years, experts expected the provision of actuarial information to raise own-longevity beliefs by approximately two years. Among profiles with large positive or negative biases, their forecasts implied an updating elasticity of roughly 0.6.