Does a worse third option change Jev's choice?

Between two gambles, does adding a third gamble that is strictly worse than one of them (the same odds, a smaller prize) make Jev pick that one more often, as it does for people?

result297 questions

Given two gambles, then a third that is a strictly worse copy of one of them, Jev leans toward the original by 5 points on average: 52% of pairs move toward it, 12% away. Jev also puts 9.7% of its weight on the worse copy itself, which no one should pick.

-0.200.000.20

shift toward the gamble with a decoy

Jev

How to read this: A single square for the decoy effect: how much Jev's share for a gamble rises when its worse copy is added, compared with when the other gamble's copy is added, with its 90% range. Zero would mean no effect.

147 gamble pairs; 90% interval on the effect [0.037, 0.065] (share of 1).

In short

A dud that no one should choose still nudges Jev toward its better twin: a small effect, but one-sided across 147 gamble pairs.
Jev also puts 9.7% of its weight on the worse copy itself, which may partly be its known trouble comparing numbers.
Whoever lays out the options can steer Jev a little without changing the real choices.

What the data shows

Gru's Plan meme: add a third gamble that's strictly worse; Jev can see it's worse; Jev drifts toward the gamble it resembles; Jev drifts toward the gamble it resembles

A small, steady decoy effect: 5 points on average toward the gamble with the worse copy (90% range roughly 4 to 7 points).
Mostly one direction: 52% of pairs move toward the target, 12% away, the rest barely move.
It doesn't always spot the dud: Jev puts 9.7% of its weight on the dominated gamble. In the example quoted under "What Jev was asked", it put 32% on gamble c, the strictly worse copy, and 46% on b.

What it means, and what it doesn't

Jev shows a mild version of the decoy effect people show: a worse option makes its neighbor more attractive. The effect is small, but it's consistent, and it comes with a real cost: weight on options no one should choose.

It isn't evidence about product menus, and part of the weight on decoys may be a numbers-reading problem rather than a preference. A version with products described in words would test the classic effect more directly.

Caveats

Gambles, not products. The classic decoy studies used products (beer, cars, restaurants) and people's taste. Here the options are money gambles written out in numbers, so the decoy has to be spotted by comparing amounts, which is closer to an arithmetic check.
Reading numbers. TypeSafe lists raw numeric values as a known weak spot for Jev. Some of the weight on the dominated gamble may be Jev failing to compare the amounts rather than being swayed by the decoy.
The project's decoys. Each decoy was built by lowering one outcome of a real gamble by 15% of its range. A bigger or more obvious gap would likely shrink both the effect and the weight on the decoy.
No human line. The human decoy effect varies a lot by setup, and these exact choices were never run with decoys on people, so there's no human number to compare with here.

Jev on this experiment

Would a person find it interesting to read?: Yes62%
Does it describe you?: Yes50%
Would you have predicted it?: No67%
How fair is the comparison?: The comparison is shaky
How much should a reader rely on it?: A little
Which caveat matters most?: Reading numbers83%

Why ask this

Cinemas sell a large popcorn by putting a medium next to it that's barely cheaper. That's the decoy effect (Huber, Payne and Puto, 1982): adding an option nobody should choose, because it's worse in every way than one of the others, makes that other option look better by comparison.

Models increasingly recommend plans, products and prices. If a model can be nudged by a dominated option, anyone laying out the menu can steer its advice without changing the real choices.

How this was done

The people and the data

Jev against itself: no people answered these three-way choices. The gambles come from choices13k (Peterson and colleagues, 2021), a large online study of how people choose between risky gambles. The project drew 150 of its pairs at random, keeping only those made of sure amounts and simple two-outcome gambles with stated odds, and for each built two decoys: one strictly worse than gamble A, one strictly worse than gamble B. Each decoy is its twin with the better outcome lowered by 15% of the gamble's range (at least $1). Three pairs couldn't take a proper decoy, so 147 pairs are analyzed.

What Jev was asked

Each pair became a three-way choice, once with A's decoy and once with B's:

Imagine you must play one of these three gambles once, for real money (wins are paid to you, losses come out of your pocket). Which do you choose: gamble_a, gamble_b or gamble_c?
gamble a: $39 with a 75% chance, or -$14 with a 25% chance · gamble b: $24 with a 90% chance, or $61 with a 10% chance · gamble c: $24 with a 90% chance, or $55 with a 10% chance

Gamble c is the decoy: the same as b but with a smaller prize. That's 297 questions (three pairs couldn't take a proper decoy), each asked with the options in shuffled orders and averaged.

How it was measured

For each pair, Jev's share for A among the two real gambles when A's decoy is present, minus the same share when B's decoy is present. Positive means the decoy helps its twin. The analysis averages over pairs with a 90% range, and also measures how much weight Jev puts on the decoy itself.

Where these questions live

297 questions across 1 topic of the map. Each opens on the map with every question in it.

Risk & Decision Style

Gamble Choices297

Every question

All 297 questions behind this result, the telling ones first: the examples the analysis points to, then the ones where Jev misses, biggest gap first.

Jev’s own answer

Showing 0 of 297