Zero-Risk Bias: Why Wiping Out One Risk Beats Shrinking Two

A risk taken to zero is worth more to people than the arithmetic says it should be. Cutting a hazard from 5 cases to 0 and cutting another from 50 to 45 prevent the same five cases, but the first closes a file and the second leaves one open, and the first is chosen and paid for more readily. The bias is not the preference for zero itself; it is the willingness to prevent fewer harms in total in order to get one count down to nothing.

Part of the pull comes from how probabilities are weighted. Prospect theory describes people as treating a move from a small chance to no chance as worth more than an equal-sized move between two middling chances, because certainty is valued out of proportion to its size. The last unit of risk removed is the one that buys the guarantee, and guarantees are priced separately from everything else.

Part of it is about attention and worry. A risk at zero can be stopped thinking about: no monitoring, no insurance, no explaining to anyone why you accepted it. A risk reduced from 50 to 30 still requires all of that. Risks are also judged one at a time rather than added up, so the question becomes whether this hazard has been dealt with rather than how many people are harmed across all of them. Zero is the only answer that reads as a complete job.

Some of the preference is defensible, and overstating the bias is the easy error here. Eliminating a hazard can be genuinely worth more than its probability suggests, because it removes the running cost of managing it and the chance of a rare catastrophic failure. The claim the research supports is narrower: when the numbers are laid out side by side and the complete option prevents fewer harms than the alternatives, a substantial minority of people still prefer it, including people whose job is to make exactly this kind of allocation.

Jonathan Baron, Rajeev Gowda and Howard Kunreuther surveyed attitudes toward hazardous waste policy and published the results in Risk Analysis in 1993. The questionnaire went to six groups with a stake in the United States Superfund program: chief executives of oil and chemical companies, economists, environmentalists, hazardous waste experts, judges and state and federal legislators. More than 400 questionnaires came back. The authors were explicit that return rates varied widely and that the samples do not represent their populations, so the useful comparison is between groups rather than to the country.

One of the cases, Case 5, was the landfill problem reproduced in the demo above, and it came in two versions given to different respondents. In the first, all three options prevented six cancer cases a year, and the option that took the smaller city to zero was simply one way of getting there: 18 percent ranked it best. In the second version, that option prevented five cases rather than six, which made it worse than both alternatives on the only measure the problem offered. It should have been ranked worst. Forty-two percent of respondents did not rank it worst, and 11 percent still ranked it best.

The bias was not spread evenly. Legislators, judges and environmentalists showed it most often; the hazardous waste experts and the economists showed it least. The authors connected the finding to the Superfund requirement for complete cleanups at individual sites, arguing that a preference people hold privately helps explain a policy that spends cleanup money where it prevents less harm.

An earlier result pointed the same way from a different direction. Viscusi, Magat and Huber studied the risk and dollar tradeoffs more than 1,500 consumers made over injuries from insecticide and toilet bowl cleaner, and reported in 1987 that the additional amount people would pay for each successive step of risk reduction diminished as the remaining risk got smaller, with a separate premium attached to eliminating one risk entirely. They found no strong evidence of an extra premium for eliminating several risks at once.

The rows below are places where one hazard can be closed out while a larger one stays open, not findings from the studies cited here.

The counter is to make the total visible, since the bias works by letting one risk be judged on its own.

• Put every option in the same unit: cases prevented, injuries avoided, expected dollars. The pull of zero fades when the three options sit in one column.
• Ask what the last unit costs. Going from 5 to 0 and from 50 to 45 prevent the same amount; if one costs five times more, that difference is the price of the guarantee, and it should be stated as such.
• Add the whole portfolio. The question is not whether this hazard is finished but how many people are harmed across all of them after the money is spent.
• Name the real benefits of closure separately. Dropped monitoring, dropped paperwork and a removed tail risk are genuine gains; write them down as line items instead of letting zero carry them silently.
• Check whether zero is really zero. Complete cleanup of one site rarely removes every path to harm, and a guarantee that still has an exclusion list is a reduction wearing a different word.
• When you present options to other people, avoid describing only one of them as eliminating a risk. That wording moves choices on its own.

Several nearby entries are about how risk is judged, and they fail in different places.

• It is not neglect of probability, where the size of a chance is ignored altogether. Zero-risk bias depends on the numbers being noticed: the appeal is specifically in the last step down to nothing.
• It is not the pseudocertainty effect, where an outcome that is certain only within one stage of a decision is treated as certain overall. There the certainty is an illusion created by the framing; here the elimination is real and still bought at too high a price.
• It is not omission bias, the preference for harm caused by doing nothing over smaller harm caused by acting. The two often appear together in safety decisions, but omission bias is about who caused the harm and zero-risk bias is about how the remaining risk is counted.