5D Library › Choice Architecture, Cumulative Prospect Theory and Computational Behavioural Economics of Pricing Interfaces · 中文
Chapter 5 Decoys and Compromises: Regularity in Discrete-Choice Models and Its Violation
Three tiers and an “unattractive” option
Many subscription and publishing products use three-tier pricing: a cheap basic tier, a middle standard tier and an expensive premium tier. A widely repeated design tip is to place next to the standard tier an option that costs more yet is no better, so that the standard tier “looks like a good deal” and sells more. Marketing and decision research call this the decoy effect, academically the asymmetric dominance or attraction effect: the added option is dominated on every attribute by the target but not by the other options. A related phenomenon is the compromise effect: an option lying “between” two others takes a larger share than in pairwise comparison.
The author’s framework asks to build a multinomial logit (MNL) and prove that it cannot produce the decoy effect, then to use Tversky and Simonson’s context-dependent model and multialternative decision field theory (MDFT) to show where the decoy and compromise effects take hold in three-tier pricing, with example prices of US$9.99, 49.90 and 149.00. This chapter follows that structure, but extends the first result to all random utility models and adds the empirical boundaries of the decoy effect.
Note
About the example prices. US$9.99, 49.90 and 149.00 are the same as 5D Library’s own prices for its e-book, paperback and hardcover omnibus editions (see the disclosure). The programs in this chapter only convert these prices into a fictitious “affordability” scale and combine it with a fictitious “quality” scale to illustrate model properties; nothing in the book shows whether this price structure works for 5D Library or any other product, and nothing validates it. The Chapter 8 experiment uses neither these prices nor the author’s products.
The multinomial logit and the independence of irrelevant alternatives
Evidence
In a random utility model option has utility and the option with the highest utility is chosen. If the are independent and identically distributed type-I extreme value, choice probabilities take the multinomial logit form (McFadden 1974). This form is equivalent to Luce’s 1959 choice axiom: the ratio of the choice probabilities of any two options does not depend on which other options are in the set — the independence of irrelevant alternatives (IIA) (Luce 1959; Train 2009).
IIA rules out the decoy effect directly. With a decoy added, the standard tier’s share becomes ; the denominator grows, so the share can only fall, and all existing options fall in the same proportion, keeping their ratios. The first part of the program below checks this.
A stronger result: regularity
IIA is a special property of the multinomial logit, and relaxing the error distribution breaks it: probit, nested logit and mixed logit all let similar options draw more share from each other. So could a more flexible random utility model produce a decoy effect? It cannot.
Evidence
Block and Marschak (1960) proved that every random utility model (whatever the distribution and correlation of the errors) satisfies regularity: if , then for any option in , (Block and Marschak 1960). The reason takes one line: the event “ has the highest utility in ” is contained in the event “ has the highest utility in ”, so its probability cannot be larger. Huber, Payne and Puto’s 1982 decoy experiments were titled precisely as “violations of regularity”: adding a dominated option raised the target’s share (Huber et al. 1982).
So the framework’s “a pure MNL cannot produce the decoy effect” is right but too narrow: no random utility model can make the target’s share rise because a decoy was added. The program below shows this with the multinomial logit, a probit with correlated errors (simulated) and a nested logit (closed form). The three options have systematic utilities basic 0, standard 0.2, premium −0.1; the decoy is 0.8 below the standard tier.
Calculation
# Multinomial logit (McFadden 1974) and regularity in random utility models.
# MNL: P(i | S) = exp(V_i) / sum_{j in S} exp(V_j). IIA: P(i)/P(j) does not depend on S.
# Regularity (Block & Marschak 1960): in ANY random utility model, P(i | S) >= P(i | T)
# whenever S is a subset of T, because {U_i = max over T} is contained in
# {U_i = max over S}. Below: (1) MNL with and without a dominated decoy; (2) a probit
# whose decoy error is strongly correlated with the target's, by simulation, and a
# nested logit in closed form: both let the decoy take share from the target (the
# similarity effect), but neither can RAISE the target's share -- an attraction effect
# is outside every random utility model.
import math, random
def mnl(V):
m = max(V.values()); e = {k: math.exp(x - m) for k, x in V.items()}
s = sum(e.values()); return {k: x / s for k, x in e.items()}
V = {"basic": 0.0, "standard": 0.2, "premium": -0.1}
dec = dict(V, decoy=V["standard"] - 0.8) # dominated by standard
p0, p1 = mnl(V), mnl(dec)
print("MNL shares basic standard premium decoy")
for lab, p in (("three tiers", p0), ("plus decoy", p1)):
print(f" {lab:15s} " + " ".join(f"{p.get(k, 0):7.3f}" for k in
("basic", "standard", "premium", "decoy")))
print(f" standard/basic ratio: {p0['standard'] / p0['basic']:.4f} ->"
f" {p1['standard'] / p1['basic']:.4f} (IIA); standard share falls by"
f" {100 * (p0['standard'] - p1['standard']):.1f} points")
def probit(rho, n=60000, with_decoy=True):
rng = random.Random(1960); wins = {k: 0 for k in dec} # common random numbers
for _ in range(n):
zs, z0 = rng.gauss(0, 1), rng.gauss(0, 1)
u = {"basic": V["basic"] + rng.gauss(0, 1),
"standard": V["standard"] + zs,
"premium": V["premium"] + rng.gauss(0, 1)}
if with_decoy:
u["decoy"] = dec["decoy"] + rho * zs + math.sqrt(1 - rho * rho) * z0
wins[max(u, key=u.get)] += 1
return {k: c / n for k, c in wins.items()}
print("probit, decoy error correlated rho with the standard tier (same 60,000 draws)")
base = probit(0.0, with_decoy=False)
print(f" no decoy: standard {base['standard']:.3f}")
for rho in (0.0, 0.5, 0.9, 0.99):
p = probit(rho)
print(f" rho {rho:4.2f}: standard {p['standard']:.3f} (change"
f" {100 * (p['standard'] - base['standard']):+5.1f} points),"
f" decoy {p['decoy']:.3f}")
def nested(mu): # standard and decoy share a nest, dissimilarity mu in (0, 1]
iv = mu * math.log(math.exp(V["standard"] / mu) + math.exp(dec["decoy"] / mu))
top = mnl({"basic": V["basic"], "premium": V["premium"], "nest": iv})
inner = mnl({"standard": V["standard"] / mu, "decoy": dec["decoy"] / mu})
return top["nest"] * inner["standard"]
print("nested logit, standard share with the decoy (no decoy: "
f"{p0['standard']:.4f})")
MUS = (1.0, 0.7, 0.4, 0.1, 0.02)
print(" " + "; ".join(f"mu {mu}: {nested(mu):.4f}" for mu in MUS))
# MNL shares basic standard premium decoy
# three tiers 0.320 0.391 0.289 0.000
# plus decoy 0.272 0.332 0.246 0.149
# standard/basic ratio: 1.2214 -> 1.2214 (IIA); standard share falls by 5.8 points
# probit, decoy error correlated rho with the standard tier (same 60,000 draws)
# no decoy: standard 0.406
# rho 0.00: standard 0.362 (change -4.4 points), decoy 0.116
# rho 0.50: standard 0.362 (change -4.3 points), decoy 0.077
# rho 0.90: standard 0.396 (change -1.0 points), decoy 0.011
# rho 0.99: standard 0.406 (change +0.0 points), decoy 0.000
# nested logit, standard share with the decoy (no decoy: 0.3907)
# mu 1.0: 0.3323; mu 0.7: 0.3318; mu 0.4: 0.3548; mu 0.1: 0.3906; mu 0.02: 0.3907In the multinomial logit the standard tier’s share falls from 0.391 to 0.332, by 5.8 percentage points, and its ratio to the basic tier stays at 1.2214 (IIA). In the probit, the more the decoy’s error is correlated with the standard tier’s, the more the decoy draws only from the standard tier: with correlation 0 the standard tier loses 4.4 points, with 0.9 only 1.0, and with 0.99 the decoy is almost never chosen and the standard tier’s share returns to 0.406. In the nested logit, as the dissimilarity parameter (the two nearly identical), the standard tier’s share approaches its no-decoy value of 0.3907 but never exceeds it; the approach is not monotone (0.3318 at is slightly below 0.3323 at ). The best these three models can do is “the decoy takes no share” (the limit of the similarity effect); they cannot make “the decoy adds share”.
This has a direct practical consequence: if an experiment reports the target tier’s share rising when a decoy is added, no random utility model — including the discrete-choice models most used in marketing analytics — can explain it; demand fitted with these models will systematically understate the decoy’s effect, and pricing optimised with them will never recommend one. To describe the attraction effect, the evaluation of an option has to depend on the other options in the set.
Tversky and Simonson’s context-dependent model
Evidence
Tversky and Simonson (1993) proposed a “context-dependent preference” model in which an option’s value has two components: an attribute weighting that reflects the background context, and binary comparisons with the other options in the current set (Tversky and Simonson 1993). The comparisons use “relative advantage”: the advantage of over sums the margins by which beats on each attribute; the disadvantage sums the margins by which trails , after an increasing convex transformation (, a loss-averse weighting of disadvantages); relative advantage is . An option’s value in set is . The model can explain both attraction and compromise effects; the latter had been documented by Simonson (1989) and by Simonson and Tversky (1992) as “extremeness aversion” (Simonson 1989; Simonson and Tversky 1992).
The program below converts the three example prices to an “affordability” scale between 0 and 1 (9.99 maps to 1.0 and 149.00 to 0.2, linear in log price), pairs it with a fictitious quality scale so that the three tiers lie exactly on a line (affordability + quality = 1.2), and no tier dominates on the simple sum of attributes. The decoy is 0.05 and 0.10 below the standard tier on the two attributes. To obtain shares the program adds a logistic error to the model values (); that is our choice, the model itself only compares values. With the model reduces to a multinomial logit.
Calculation
# Tversky & Simonson (1993) context-dependent model on a FICTITIOUS three-tier page.
# Attribute values v1 = affordability, v2 = quality. Relative advantage of x over y:
# A(x, y) = sum_i max(v_i(x) - v_i(y), 0)
# D(x, y) = sum_i delta(max(v_i(y) - v_i(x), 0))
# R(x, y) = A / (A + D); V(x | S) = v1(x) + v2(x) + theta * sum_{y in S} R(x, y)
# delta(t) = t + kappa t^2 is increasing and convex with delta(t) >= t (loss aversion).
# To get shares we add a logit error, P(x | S) ~ exp(V(x | S) / tau) (our choice; the
# model itself compares values). theta = 0 reduces it to a multinomial logit.
import math
def afford(price): # log-price scale between 9.99 (1.0) and 149.00 (0.2)
return 1 - 0.8 * math.log(price / 9.99) / math.log(149.0 / 9.99)
S_ = afford(49.90) # the three tiers lie on one line: v1 + v2 = 1.2
TIERS = {"basic": (afford(9.99), 0.20), "standard": (S_, 1.2 - S_),
"premium": (afford(149.0), 1.00)}
DECOY = {"decoy": (S_ - 0.05, 1.1 - S_)} # worse than standard on both attributes
def rel(x, y, kappa):
a = sum(max(p - q, 0) for p, q in zip(x, y))
d = sum(t + kappa * t * t for t in (max(q - p, 0) for p, q in zip(x, y)))
return 0.5 if a + d == 0 else a / (a + d)
def shares(opts, theta, kappa=1.0, tau=0.1):
V = {k: sum(x) + theta * sum(rel(x, y, kappa) for j, y in opts.items() if j != k)
for k, x in opts.items()}
m = max(V.values()); e = {k: math.exp((v - m) / tau) for k, v in V.items()}
s = sum(e.values()); return {k: x / s for k, x in e.items()}
print("attribute values (affordability, quality):")
ALL = {**TIERS, **DECOY}
print(" " + "; ".join(f"{k} {x[0]:.3f}, {x[1]:.2f}" for k, x in ALL.items()))
print("attraction: standard's share, three tiers -> three tiers plus decoy")
print("compromise: standard/basic odds, {basic, standard} -> {basic, standard, premium}")
print(" theta kappa 3 tiers +decoy change odds 2 tiers odds 3 tiers")
for theta in (0.0, 0.01, 0.03, 0.1, 0.3, 0.6):
for kappa in ((0.0, 2.0) if theta else (0.0,)):
s3 = shares(TIERS, theta, kappa)
s4 = shares({**TIERS, **DECOY}, theta, kappa)
two = {k: TIERS[k] for k in ("basic", "standard")}
s2 = shares(two, theta, kappa)
print(f" {theta:5.2f} {kappa:5.1f} {s3['standard']:8.3f}"
f" {s4['standard']:6.3f}"
f" {100 * (s4['standard'] - s3['standard']):+6.1f}"
f" {s2['standard'] / s2['basic']:12.3f}"
f" {s3['standard'] / s3['basic']:12.3f}")
s4 = shares({**TIERS, **DECOY}, 0.3, 2.0)
print(f"theta 0.3, kappa 2: decoy share {s4['decoy']:.3f}; shares"
f" basic {s4['basic']:.3f}, premium {s4['premium']:.3f}")
# attribute values (affordability, quality):
# basic 1.000, 0.20; standard 0.524, 0.68; premium 0.200, 1.00; decoy 0.474, 0.58
# attraction: standard's share, three tiers -> three tiers plus decoy
# compromise: standard/basic odds, {basic, standard} -> {basic, standard, premium}
# theta kappa 3 tiers +decoy change odds 2 tiers odds 3 tiers
# 0.00 0.0 0.333 0.310 -2.3 1.000 1.000
# 0.01 0.0 0.333 0.321 -1.3 1.000 1.000
# 0.01 2.0 0.335 0.325 -1.0 1.000 1.010
# 0.03 0.0 0.333 0.342 +0.8 1.000 1.000
# 0.03 2.0 0.339 0.354 +1.6 1.000 1.030
# 0.10 0.0 0.333 0.416 +8.3 1.000 1.000
# 0.10 2.0 0.351 0.464 +11.2 1.000 1.105
# 0.30 0.0 0.333 0.625 +29.1 1.000 1.000
# 0.30 2.0 0.389 0.751 +36.2 1.000 1.349
# 0.60 0.0 0.333 0.850 +51.6 1.000 1.000
# 0.60 2.0 0.446 0.949 +50.3 1.000 1.821
# theta 0.3, kappa 2: decoy share 0.005; shares basic 0.105, premium 0.140With , adding the decoy lowers the standard tier’s share by 2.3 points, as regularity requires. For small it still falls (: −1.3 or −1.0 points): the dilution by the decoy (the multinomial-logit part) outweighs the gain from relative advantage; at it turns positive (+0.8 or +1.6), so in this example () the threshold for the attraction effect lies between 0.01 and 0.03, its exact position depending on and on the decoy’s own share. As grows, the attraction effect strengthens quickly: at the standard tier gains 8.3 points () or 11.2 points (), at it gains 29.1 or 36.2, while almost nobody chooses the decoy itself (0.005). The compromise effect appears only with a convex disadvantage function (): the standard tier’s odds against basic rise, when premium is added, from 1.000 to 1.105 (), 1.349 () and 1.821 (); with the three tiers are symmetric on the line, relative advantages cancel and the odds stay at 1.000. In this model the attraction effect needs “pairwise comparisons enter the evaluation” with enough weight; the compromise effect also needs “disadvantages weigh more than advantages”.
Multialternative decision field theory
Evidence
Roe, Busemeyer and Townsend’s (2001) multialternative decision field theory (MDFT) treats choice as preferences accumulating over time: at each moment attention falls at random on one attribute, each option’s value on that attribute is compared with the mean of the others to give a momentary “valence”, and the preference state evolves as , where the diagonal of the feedback matrix is memory and the off-diagonal entries are lateral inhibition between options, weakening as the distance between two options in attribute space grows. The paper reports that setting lateral inhibition to zero removes the attraction effect: the dominated decoy accumulates negative preference which, through its strong negative link to the target, “two negatives make a positive”, pushes the target up; the compromise effect likewise depends on lateral inhibition, which makes the middle option negatively correlated with both extremes and gives it an edge in the comparison (Roe et al. 2001).
The program below uses the same four options as the previous section, with lateral inhibition ( the Euclidean distance between options; the exponential form is one common choice, the paper only requires inhibition to fall with distance); each simulated chooser accumulates 100 steps and picks the highest preference.
Calculation
# Multialternative decision field theory (Roe, Busemeyer & Townsend 2001) on the same
# FICTITIOUS tiers. Preferences evolve as P(t+1) = S P(t) + C M W(t+1) + e(t+1):
# M = options x attributes, W(t) attends to ONE attribute per step (probability 1/2 each),
# C subtracts the mean of the other options, e ~ N(0, sig^2). Feedback S has self-memory
# 1 - phi1 on the diagonal and lateral inhibition -phi2 * exp(-d_ij^2 / ell) off it,
# d = distance between options in attribute space (the distance form is one common
# choice; the 2001 paper only requires inhibition to fall with distance). Choice is the
# largest P after T steps. phi2 = 0 switches lateral inhibition off.
import math, random
def afford(price):
return 1 - 0.8 * math.log(price / 9.99) / math.log(149.0 / 9.99)
S_ = afford(49.90) # the tiers lie on one line: afford + quality = 1.2
OPT = {"basic": (afford(9.99), 0.20), "standard": (S_, 1.2 - S_),
"premium": (afford(149.0), 1.00), "decoy": (S_ - 0.05, 1.1 - S_)}
def mdft(names, phi2, n=1500, T=100, phi1=0.01, ell=0.2, sig=0.1, seed=2001):
rng = random.Random(seed); m = len(names); X = [OPT[k] for k in names]
S = [[(1 - phi1) if i == j else
-phi2 * math.exp(-sum((a - b) ** 2 for a, b in zip(X[i], X[j])) / ell)
for j in range(m)] for i in range(m)]
wins = [0] * m
for _ in range(n):
P = [0.0] * m
for _ in range(T):
a = 0 if rng.random() < 0.5 else 1
val = [x[a] for x in X]; tot = sum(val)
V = [val[i] - (tot - val[i]) / (m - 1) + rng.gauss(0, sig) for i in range(m)]
P = [sum(S[i][j] * P[j] for j in range(m)) + V[i] for i in range(m)]
wins[max(range(m), key=P.__getitem__)] += 1
return {k: c / n for k, c in zip(names, wins)}
three, four = ["basic", "standard", "premium"], ["basic", "standard", "premium", "decoy"]
print("MDFT shares (1,500 simulated choosers per set, common seed, T = 100 steps)")
print("narrow inhibition: phi1 = 0.01, ell = 0.2")
print(" phi2 set basic standard premium decoy")
res = {}
for phi2 in (0.0, 0.01, 0.02, 0.03):
for lab, names in (("2 tiers", ["basic", "standard"]), ("3 tiers", three),
("3 tiers + decoy", four)):
p = mdft(names, phi2); res[phi2, lab] = p
print(f" {phi2:4.2f} {lab:15s} " + " ".join(
f"{p[k]:7.3f}" if k in p else " -" for k in four))
for phi2 in (0.0, 0.01, 0.02, 0.03):
a = res[phi2, "3 tiers + decoy"]["standard"] - res[phi2, "3 tiers"]["standard"]
o2 = res[phi2, "2 tiers"]["standard"] / res[phi2, "2 tiers"]["basic"]
o3 = res[phi2, "3 tiers"]["standard"] / res[phi2, "3 tiers"]["basic"]
print(f" phi2 {phi2:4.2f}: attraction {100 * a:+5.1f} points;"
f" standard/basic odds {o2:.2f} -> {o3:.2f} with premium")
print("broad inhibition and faster decay: phi1 = 0.06, ell = 1.0")
for phi2 in (0.05, 0.10):
r = {lab: mdft(names, phi2, phi1=0.06, ell=1.0) for lab, names in
(("2", ["basic", "standard"]), ("3", three), ("4", four))}
print(f" phi2 {phi2:4.2f}: standard {r['2']['standard']:.3f} (2 tiers),"
f" {r['3']['standard']:.3f} (3), {r['4']['standard']:.3f} (+decoy);"
f" odds {r['2']['standard'] / r['2']['basic']:.2f} ->"
f" {r['3']['standard'] / r['3']['basic']:.2f}")
# MDFT shares (1,500 simulated choosers per set, common seed, T = 100 steps)
# narrow inhibition: phi1 = 0.01, ell = 0.2
# phi2 set basic standard premium decoy
# 0.00 2 tiers 0.511 0.489 - -
# 0.00 3 tiers 0.476 0.063 0.461 -
# 0.00 3 tiers + decoy 0.503 0.070 0.427 0.000
# 0.01 2 tiers 0.507 0.493 - -
# 0.01 3 tiers 0.473 0.071 0.456 -
# 0.01 3 tiers + decoy 0.357 0.355 0.288 0.000
# 0.02 2 tiers 0.508 0.492 - -
# 0.02 3 tiers 0.465 0.083 0.453 -
# 0.02 3 tiers + decoy 0.145 0.772 0.083 0.000
# 0.03 2 tiers 0.513 0.487 - -
# 0.03 3 tiers 0.453 0.103 0.445 -
# 0.03 3 tiers + decoy 0.012 0.981 0.007 0.000
# phi2 0.00: attraction +0.7 points; standard/basic odds 0.96 -> 0.13 with premium
# phi2 0.01: attraction +28.3 points; standard/basic odds 0.97 -> 0.15 with premium
# phi2 0.02: attraction +68.9 points; standard/basic odds 0.97 -> 0.18 with premium
# phi2 0.03: attraction +87.9 points; standard/basic odds 0.95 -> 0.23 with premium
# broad inhibition and faster decay: phi1 = 0.06, ell = 1.0
# phi2 0.05: standard 0.509 (2 tiers), 0.107 (3), 0.327 (+decoy); odds 1.04 -> 0.23
# phi2 0.10: standard 0.483 (2 tiers), 0.436 (3), 0.993 (+decoy); odds 0.94 -> 1.83With narrow inhibition () and no lateral inhibition (), adding the decoy changes the standard tier’s share by 0.7 points, which reflects simulation noise and the change of valence scale when the contrast matrix averages over 3 other options instead of 2; with the share rises by 28.3 points and with by 87.9 points. But the same parameters give no compromise effect: when premium is added, the standard tier’s odds against basic fall from about 0.96 to between 0.13 and 0.23. The reason is that attention falls on one attribute at a time and the middle option is best on neither, which puts it at a disadvantage in the random walk; only with broader, stronger inhibition and faster decay (, , ) do the odds rise from 0.94 to 1.83, a compromise effect — by which point adding the decoy lifts the standard tier’s share from 0.436 to 0.993 (an attraction effect of about +55.7 points), so that nearly everyone chooses it. Figure 4 plots the attraction effect against the parameters of both models.
The common lesson of the two models is that the attraction effect is mechanically easy to produce, as long as an option’s evaluation depends strongly enough on comparisons with the others; the compromise effect needs more special conditions. But “mechanically easy” does not mean “bound to appear on a real page”; that is for the evidence.
Which model to use when analysing page tests
These results have a direct consequence for analysing page-test data. If a randomised test compares pages with and without a decoy, the simplest and least model-dependent analysis compares the shares choosing the target tier in the two arms directly (as Chapter 8 does); no choice model is needed. If a model is used to carry the result over to other price combinations, random utility models such as the multinomial and nested logit structurally cannot reproduce the attraction effect, and their fit will read the decoy’s effect as zero or negative; one needs Tversky and Simonson’s model, MDFT or another model that allows context dependence. Rieskamp, Busemeyer and Mellers’s 2006 review brings together the evidence on violations of “rational choice principles” such as regularity and the various forms of stochastic transitivity, compares the kinds of model that can explain them, and discusses how data can discriminate between those models (Rieskamp et al. 2006). For this book the point is that the choice of model is not a technicality: it decides in advance whether the analysis can “see” a decoy effect at all.
Where the decoy effect takes hold
Evidence
Frederick, Lee and Baskin (2014) reported 38 studies: in abstract choices with numerically presented attributes (gambles with numerical probabilities, televisions with numerical quality ratings) the attraction effect replicated (target shares rising from 14 % to 28 %, from 21 % to 37 %, from 33 % to 57 %); but in 27 studies with at least one attribute presented perceptually or experienced (apartment pictures, juice, mints, bottled water and others), none showed a significant attraction effect, and bottled water showed a significant reverse “repulsion effect” (target share falling from 70 % to 52 %) (Frederick et al. 2014). Yang and Lynn (2014) reported similar failures (Yang and Lynn 2014). In reply, the original authors Huber, Payne and Puto accepted that the effect has boundaries and listed the conditions where it is stronger: weak prior preferences, dominance visible at a glance (as with numerical attributes), a rarely chosen decoy, core options with similar shares and people with similar values; it is weaker or absent with rich or sensory attributes, strong prior preferences, or a decoy that some people find attractive (Huber et al. 2014). Heath and Chatterjee’s 1995 meta-analysis found that decoys reduce the shares of lower-quality competitors more than those of higher-quality ones, and violate regularity only when a higher-quality brand is the target (Heath and Chatterjee 1995). Trueblood and co-authors (2013) observed all three context effects in perceptual judgements (comparing rectangle areas) (Trueblood et al. 2013), but Frederick and co-authors failed to replicate the attraction effect there; Spektor, Bhatia and Gluth’s 2021 review discusses how hard it is to obtain context effects reliably across tasks and presentations (Spektor et al. 2021). In the field, Wu and Cosguner (2020) analysed diamond sales at a large online jewellery retailer and estimated that only 11 % to 25 % of shoppers noticed the dominance relations; among those who did, the sales hazard of dominant diamonds rose 1.8 to 3.2 times, and the decoy effect raised the retailer’s gross profit by 14.3 % (Wu and Cosguner 2020).
Taken together, the boundaries of the decoy effect in three-tier pricing can be stated as four points, all from the evidence above rather than from the models:
- Dominance has to be seen. Numerical, side-by-side attributes in the same units favour the effect; pictures, taste, experiential attributes or attributes that need calculation to compare work against it. In field data only 11 % to 25 % noticed.
- The decoy itself should hardly ever be chosen. If a sizeable share chooses it, it is not dominated for some people or the dominance was not seen; the Chapter 8 experiment treats “decoy chosen by more than 10 %” as a threshold for non-interpretability.
- The core options must not be too far apart. If people already strongly prefer one tier, a decoy can hardly change their choice.
- The direction can reverse. Repulsion effects have been observed with perceptual attributes; a page designer cannot assume the direction in advance.
Hypothesis · not established · The decoy effect varies with the visibility of dominance
The author’s hypothesis (not established): on a three-tier subscription pricing page, the decoy effect increases monotonically with the visibility of the dominance relation: it is largest when the difference between decoy and target is listed side by side in numbers, smaller when described in words and close to zero when shown with icons or pictures. What would show it wrong: in a pre-registered experiment randomising “decoy or not × three presentation formats”, the decoy effect shows no monotone trend across formats (the confidence interval for a trend test includes zero).
Chapter summary
The multinomial logit satisfies IIA, so adding a decoy lowers the shares of all existing options proportionally. More generally, every random utility model satisfies regularity, so the decoy effect violates all random utility models; correlated errors can at most stop the decoy from taking share, never make the target rise. In the Tversky–Simonson model the attraction effect appears once the weight on relative advantage passes a threshold (between 0.01 and 0.03 in the example), while the compromise effect also needs a convex disadvantage function; in MDFT both depend on lateral inhibition, the compromise effect needing broader and stronger inhibition. Empirically the attraction effect replicates in abstract choices with numerical attributes but often vanishes or even reverses with perceptual and experiential attributes, and only 11 % to 25 % notice dominance in the field. The example prices are the same as 5D Library’s own, and nothing in this chapter validates them.
References
Fan, Y. (2026). Choice Architecture, Cumulative Prospect Theory and Computational Behavioural Economics of Pricing Interfaces: From Loss-Aversion Value Functions and Asymmetric Dominance to Hyperbolic Discounting, Read Beside the Qingzhong Chapters of the Guanzi [选择架构、累积前景理论与销售界面价格锚定的计算行为学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23836260 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).