Polarizing Time Preferences: the Ambivalent Role of Uncertainty in Savings

The information treatments directly target subjective uncertainty about respondents’ own future idiosyncratic income. They are delivered as AI generated videos with embedded subtitles. Screenshots and full scripts are provided in the Appendix.

Baseline treatment.
Respondents randomly assigned to the first treatment view a video that presents quantitative and qualitative evidence that households systematically make large forecast errors when predicting income two years ahead.

First, quantitative information about the true forecast error distribution is provided, using Rozsypal and Schlafmann (2023) data on income forecast errors. The video reports that the mean absolute forecast error is 29 percentage points, which is roughly £10,000 on a £35,000 annual salary.

Then, the video highlights qualitative reasons why large errors might be happening over a two year horizon, including unplanned promotions, job loss, switches of employer or career and personal changes.

The treatment is designed to update beliefs about dispersion rather than levels. To avoid directly shifting mean expectations, the script emphasizes throughout that forecast errors can be positive or negative with roughly equal likelihood. It states that about half of people overestimate and about half underestimate their future income and later reiterates that a respondent is just as likely to be ``too optimistic" as ``too pessimistic". A comprehension check immediately after the video asks, “When predicting their income two years from now, people tend to:” with the correct choice “Be just as likely to overestimate as to underestimate.”

Alternative upside treatment.
The second treatment, assigned to a smaller group and used primarily as a robustness check, presents the same quantitative content as the baseline treatment but shifts the emphasis to upside risk in the qualitative part. The script states that, on average, individuals underestimate how much their income will grow and it specifically highlights reasons for positive surprises such as unplanned promotions, or job and career switches.

2025-08-12

2025-09-08

Discount factor and precautionary savings

Discount factors estimated using Convex time budgets method from Andreoni and Sprenger (2012)

The questionnaire ran in Qualtrics with recruitment by Bilendi. Eligibility required residence in the United Kingdom, age between 20 and 55 years and current employment in a paid job. The final sample includes 2,000 respondents, with a median completion time of 17 minutes. Respondents first provided informed consent.

The pre-treatment block elicits time preferences, using a CTB task that presents 10 allocation questions between a sooner payoff at seven days and a later payoff at two years. It then measures income expectations and uncertainty for one year ahead horizon. It also collects socio demographic variables as well as a self-reported measure of liquid savings.

Then respondents assigned to the treatment arms then viewed a video of about 1.30 minutes. The two treatment branches used slightly different versions of the video, but both aim at shifting uncertainty by presenting evidence on the typical magnitude of forecast errors and outlined common sources of error when individuals predict their own income two years ahead. I required respondents to watch the video until the end and logged video watch time to detect attempts to speed through the content. Immediately after the video, respondents answered attention check questions that verified comprehension of the content. The control arm proceeded without the video and without these follow up questions. All other survey content was identical across arms.

After the treatment block, income uncertainty is elicited for two year ahead income using the same method. The CTB time preference task was repeated with the same interest rates but with a different presentation so that respondents don't feel like they are asked the same questions twice. Additional outcomes included self-reported anxiety and cognitive difficulty of thinking ahead, as well as precautionary savings and attitudes to uncertainty.

After the pre-treatment block, I randomly assigned respondents to one of three arms: control with 840 respondents (42% of sample), baseline treatment with 837 respondents (42%) and alternative treatment with 323 respondents (16%).

Individual

No

2,000

2,000 individuals

840 respondents in control, 1,160 treated (837 in baseline treatment arm, 323 in upside treatment arm)