Omission Bias: Why Doing Nothing Feels Less Wrong Than Acting

The arithmetic in the demo above is not close. A disease that kills 10 children in 10,000 and a vaccine that kills 5 leaves 5 children alive per 10,000 who would otherwise have died, and anyone counting deaths would vaccinate. What stops some people is that the 5 deaths from the vaccine would be caused by a decision, while the 10 deaths from the flu would merely be allowed. The term has a second, unrelated use in media criticism, where bias by omission means a news outlet leaving out facts that would change the picture; this page is about the psychological finding.

In most of ordinary life the distinction between doing and allowing is worth keeping. Omissions often come from ignorance and commissions rarely do, and commissions usually signal a stronger intention. Mark Spranca, Elisa Minsk and Jonathan Baron argued in 1991 that the bias is the overextension of that useful distinction to cases stripped of everything that made it useful: the agent knows the consequences of both options, intends the same thing in both, and has to pick one.

When people are asked why, two answers come back. Some say an omission is not a cause, so the deaths that follow from it are not theirs. Others do not hold themselves responsible for an outcome that would have happened had they never been consulted, even though they were consulted and did know.

Two things are worth saying about the size of this. First, the pull is stronger when something is unknown. In Ritov and Baron's experiments, telling people that a risk group for vaccine deaths existed, while holding the overall death rate constant and making clear that no test for membership could be given, made them less willing to vaccinate: 18 subjects moved that way against 3 who moved the other way. A risk group for the flu did nothing. Second, most people in these studies show no bias at all; Ritov and Baron reported that a substantial minority did, which is enough to shape a program or a law.

Ilana Ritov and Jonathan Baron published "Reluctance to vaccinate: Omission bias and ambiguity" in the Journal of Behavioral Decision Making in 1990. In the first experiment, 53 undergraduates were told that a disease kills 10 of every 10,000 children and that a vaccine costing 2 dollars a child prevents it completely but has side effects that kill some children. They were handed a table showing each possible side-effect death rate from 0 to 9 per 10,000 with the net decrease in deaths and the cost per life saved, and asked for the maximum risk the government should tolerate in a compulsory program.

In the basic version, 23 percent said no risk at all should be tolerated, 57 percent of answers fell between 1 and 8 per 10,000, and 9 percent gave 9 or 10, the answers that save the most lives. A second version told them that the children susceptible to the flu were the same ones susceptible to the vaccine, so that nobody would die who would not have died anyway: 47 percent then chose the maximum risk, against 9 percent in the basic version. Raising the price fivefold changed almost nothing, which rules out cost as the explanation.

The second experiment put the decision to 28 students, of whom 20 gave usable answers, and asked for the highest vaccine death rate at which they would still vaccinate, once as a parent and once as a voter on a law. In the basic case the mean was 5.458 per 10,000 for the personal decision and 5.750 for the law, against a flu killing 10 per 10,000, so the average participant demanded that the vaccine be nearly twice as safe as the disease. Mentioning a risk group for the vaccine lowered the tolerated rate to 4.350, and describing the same number of deaths as the vaccine failing rather than as side effects raised it to 7.125.

That last comparison is the clearest statement of what the bias is about: the number of children who die is identical, and what changes is whether the vaccine caused the deaths or failed to prevent them. Ritov and Baron separated this from a preference for the status quo in a 1992 paper and found the act-omission distinction was doing the work. The demo above moves both at once, since one version makes vaccination the action and the other makes it the default.

The pattern appears wherever the safer option on paper requires somebody to act. The rows below are illustrations, not findings from the studies cited on this page.

The bias lives in the difference between a choice and a non-choice, so the counter is to make both options choices.

• Write the two options as two actions. "Vaccinate 10,000 children" and "decline to vaccinate 10,000 children" are both decisions with a number of deaths attached.
• Count the harm on both sides. The question is not whether the vaccine kills anyone but whether more people die with it or without it.
• Run the reversal test. If the current arrangement were the one you had to vote in rather than the one you had to vote out, would you still choose it? That is the comparison the demo above makes across visitors.
• Watch for missing information. In Ritov and Baron's experiments, naming a risk group nobody could test for made people less willing to act even with the total risk held constant. Information you cannot get is not a reason to wait for it.
• Separate the moral question from the arithmetic. Some people hold, on reflection, that causing a death is worse than allowing one. That is a position to defend in the open, not a reason to avoid doing the sum.

Omission bias is often lumped together with its neighbors, and the differences come down to what is being held constant.

• It is not status quo bias, the preference for leaving an arrangement in place. The two usually coincide, because doing nothing normally preserves the status quo, but Ritov and Baron separated them in 1992 and found that the act-omission distinction mattered more than the existing arrangement did.
• It is not loss aversion, the finding that a loss weighs more than an equal gain. Loss aversion is part of why harm from an omission reads as a benefit not collected, but omission bias concerns who is responsible for a harm rather than how the harm is scaled.
• It is not zero-risk bias, the preference for eliminating one risk completely over cutting a larger risk in part. A person showing zero-risk bias is willing to act, but on the wrong target.