5D Library › Choice Architecture, Cumulative Prospect Theory and Computational Behavioural Economics of Pricing Interfaces · 中文
Chapter 4 Loss Aversion and Framing: From Individual Choice to a “Conversion Gap”
“Losses weigh 2.25 times as much as gains”
The author’s framework wants to use cumulative prospect theory to “quantify the conversion gap between loss and gain frames on a checkout page”. Behind this lies a common argument: since people feel losses 2.25 times as strongly as gains, describing an offer as “lose it if you don’t buy” should work far better than “get it if you buy”. This chapter checks the argument in two steps. First, what is, where it comes from and how stable it is. Second, even with known, what is missing between an individual value function and a “conversion rate” on a page.
Evidence on the loss-aversion coefficient
Evidence
Brown, Imai, Vieider and Camerer (2024) meta-analysed 607 estimates of the loss-aversion coefficient from 150 articles published between 1992 and 2017 in economics, psychology, neuroscience and other fields; the mean is 1.955, with a 95 % interval for the true mean of 1.820 to 2.102, and the study-design features they recorded explain only a small part of the variation between estimates (Brown et al. 2024). In random utility models of brand choice, Hardie, Johnson and Fader (1993) estimated a multi-attribute reference-dependent multinomial logit on scanner data and found significant loss aversion (Hardie et al. 1993); Neumann and Böckenholt (2014), pooling 109 observations from 33 studies, found loss aversion in product choice but with substantial variation across research contexts, products, consumers and model specifications (especially the latter) (Neumann and Böckenholt 2014). On the other side, Gal and Rucker (2018) reviewed the evidence and concluded that it does not support the view that losses are, on balance, more impactful than gains, and discussed why the belief persists in the social sciences (Gal and Rucker 2018).
Put together, these do not say that “loss aversion does not exist”; they say that a single that can be applied anywhere does not exist. The estimate depends on the task (risky choice, riskless exchange, brand choice), on the stakes and — as the program below shows — on the model assumed when estimating it.
The coefficient depends on the model
Consider a fictitious decision-maker whose true preferences are: gain exponent , loss exponent , , and 1992 weights on probability one-half (, ). She faces a series of gambles “win with probability 1/2, lose with probability 1/2” and accepts or rejects them with logistic noise. An analyst estimates her loss aversion with the most common simplified model: linear value and equal weights, .
Calculation
# What a "loss-aversion coefficient" measures depends on the model it is estimated in.
# A FICTITIOUS chooser accepts or rejects 50-50 mixed gambles (win G, lose L). Her true
# CPT parameters: alpha = 0.88 (gains), beta = 0.80 (losses), lambda = 1.5, TK (1992)
# weights w+(.5) = 0.421, w-(.5) = 0.454. An analyst fits lambda by maximum likelihood
# assuming linear value and equal weights: P(accept) = logistic((G - lam L) / s).
# Same data, different stake ranges -> different "lambda".
import math, random
AL, BE, LAM, WP, WM = 0.88, 0.80, 1.5, 0.421, 0.454
def accept_prob(G, L, tau=0.5):
u = WP * G ** AL - WM * LAM * L ** BE
return 1 / (1 + math.exp(-u / tau))
def fit(data):
best = None
for i in range(50, 401): # lam 0.50 .. 4.00
lam = i / 100
for s in (0.25, 0.5, 1, 2, 4, 8, 16, 32, 64):
ll = 0.0
for G, L, y in data:
z = (G - lam * L) / s * (1 if y else -1) # stable log-likelihood
ll -= math.log1p(math.exp(-z)) if z > 0 else math.log1p(math.exp(z)) - z
if best is None or ll > best[0]:
best = (ll, lam, s)
return best[1]
rng = random.Random(1992)
print("true: alpha 0.88, beta 0.80, lambda 1.5, w+(.5) .421, w-(.5) .454")
print("indifference loss L* for gain G (P(accept) = 1/2), and the ratio G / L*:")
for G in (10, 100, 1000):
Ls = ((WP * G ** AL) / (WM * LAM)) ** (1 / BE)
print(f" G = {G:5d}: L* = {Ls:7.2f}, G / L* = {G / Ls:.3f}")
print("analyst's linear-value lambda, 400 simulated choices per stake range")
for lab, lo, hi in (("small stakes, G in 2..20", 2, 20),
("medium, G in 20..200", 20, 200),
("large, G in 200..2000", 200, 2000)):
data = []
for _ in range(400):
G = rng.uniform(lo, hi); L = G * rng.uniform(0.2, 1.2)
data.append((G, L, rng.random() < accept_prob(G, L)))
print(f" {lab:26s}: lambda-hat = {fit(data):.2f}")
print(f"true lambda is {LAM}; only if alpha = beta and w+ = w- would G / L* equal")
print(f" lambda^(1/alpha) = {LAM ** (1 / AL):.3f} at every stake")
# true: alpha 0.88, beta 0.80, lambda 1.5, w+(.5) .421, w-(.5) .454
# indifference loss L* for gain G (P(accept) = 1/2), and the ratio G / L*:
# G = 10: L* = 6.90, G / L* = 1.449
# G = 100: L* = 86.88, G / L* = 1.151
# G = 1000: L* = 1093.75, G / L* = 0.914
# analyst's linear-value lambda, 400 simulated choices per stake range
# small stakes, G in 2..20 : lambda-hat = 1.43
# medium, G in 20..200 : lambda-hat = 1.13
# large, G in 200..2000 : lambda-hat = 0.88
# true lambda is 1.5; only if alpha = beta and w+ = w- would G / L* equal
# lambda^(1/alpha) = 1.585 at every stakeFor the true decision-maker, the ratio of gain to the indifference loss is 1.449 at , 1.151 at and 0.914 at : the larger the stakes, the more willing she is, because the loss exponent 0.80 is smaller than the gain exponent 0.88. From the same person’s choices, the analyst’s linear model estimates a “” of 1.43, 1.13 and 0.88 for small, medium and large stakes; the true value 1.5 never appears. Only if and the weights are equal is the ratio equal to at every stake. The “loss-aversion coefficient” is thus a parameter within a model: values of from studies using different specifications and stake ranges should not be expected to agree, still less should any one number be treated as a constant of human nature for page design.
Framing effects: reliable, but of limited size
Evidence
In Tversky and Kahneman’s 1981 “Asian disease problem”, 600 people face a disease. In the gain frame, between programme A, “200 people will be saved”, and programme B, “a 1/3 probability that 600 will be saved and 2/3 that nobody will be saved”, 72 % of participants () chose A; in the loss frame, between C, “400 people will die”, and D, “a 1/3 probability that nobody will die and 2/3 that 600 will die”, only 22 % () chose C (Tversky and Kahneman 1981). Kühberger’s 1998 meta-analysis of 136 empirical papers with nearly 30,000 participants and 230 effect sizes judged framing to be a reliable phenomenon with an overall effect “of small to moderate size”; effect sizes depended markedly on how framing was operationalised (changing reference points or the salience of outcomes) and on the response mode (choice or rating) (Kühberger 1998).
Reference-dependent choice models and the choice set
Loss aversion can also be put directly into the discrete-choice models of Chapter 5. Hardie and co-authors gave each brand “gains” or “losses” in price and quality relative to a reference point (for example the brand last bought), with a larger coefficient on losses than on gains, and entered these terms in the systematic utility of a multinomial logit (Hardie et al. 1993). With the reference point given, such a model is still a random utility model and satisfies the regularity of Chapter 5: the reference point does not move with the options on the page, so adding an option cannot raise the share of an existing one.
Things change when the reference point is set by the choice set itself. If customers take some “middle” of the options on a page as their reference point, adding an option moves the reference point and so changes the gains and losses of the others; loss aversion can then produce set-dependent choice. The convex weighting of disadvantages in Tversky and Simonson’s 1993 model, and the “extremeness aversion” of Simonson and Tversky (1992), can be read along these lines: each option’s disadvantages relative to the others weigh more than its advantages, and an option in the middle has smaller disadvantages in both directions (Tversky and Simonson 1993; Simonson and Tversky 1992). In this book’s structure loss aversion is thus both a parameter of the individual value function in Chapters 3 and 4 and one source of the compromise effect in Chapter 5; the idea is the same, the reference point different.
The missing link function
Cumulative prospect theory gives each option a value ; to get a share (how many choose it) one has to assume randomness between people, or within one person over time. The usual device is a logistic link: , with scale parameter . This is not part of cumulative prospect theory and must be estimated separately. The program below first puts the 1992 median parameters into the Asian disease problem, then turns to a fictitious checkout page.
Calculation
# From CPT values to a "conversion rate" needs a link that CPT does not supply.
# Part A: Tversky & Kahneman's (1981) disease problem under the 1992 medians. Gain frame:
# A = 200 saved for sure, B = 600 saved with p 1/3. Loss frame: C = 400 die for sure,
# D = nobody dies with p 1/3, 600 die with p 2/3. Reference point: the status quo of each
# frame. P(sure option) = 1 / (1 + exp(-(V_sure - V_risky) / tau)) for a logit scale tau.
# Part B (illustrative checkout): a 10-unit discount described as a gain ("save 10") or
# as a loss ("lose your 10 if you leave"). The utility shift is v(10) or lam * v(10);
# conversion = logistic(base logit + shift / tau). The base rate is set to 30 %.
import math
A, LAM, GG, GL = 0.88, 2.25, 0.61, 0.69
def w(p, g):
return p ** g / (p ** g + (1 - p) ** g) ** (1 / g)
def v(x):
return x ** A if x >= 0 else -LAM * (-x) ** A
def logistic(z):
return 1 / (1 + math.exp(-z))
VA, VB = v(200), w(1 / 3, GG) * v(600)
VC, VD = v(-400), w(2 / 3, GL) * v(-600)
print(f"gain frame: V(A) = {VA:.1f}, V(B) = {VB:.1f}; loss frame: V(C) = {VC:.1f},"
f" V(D) = {VD:.1f}")
print("published share choosing the sure option (TK 1981): 72 % of 152 (A),"
" 22 % of 155 (C)")
print(" tau P(A) P(C) frame effect (points)")
for tau in (1, 5, 10, 20, 50, 100, 1000):
pa, pc = logistic((VA - VB) / tau), logistic((VC - VD) / tau)
print(f" {tau:4d} {pa:5.2f} {pc:5.2f} {100 * (pa - pc):6.1f}")
lo, hi = 1.0, 1000.0
for _ in range(80): # tau that reproduces P(A) = 0.72
mid = (lo + hi) / 2
lo, hi = (mid, hi) if logistic((VA - VB) / mid) > 0.72 else (lo, mid)
print(f" tau fitted to P(A) = 0.72: {lo:.1f}; it then predicts P(C) ="
f" {logistic((VC - VD) / lo):.2f} (observed 0.22)")
print()
b0 = math.log(0.3 / 0.7)
print("checkout (illustrative): base conversion 30 %, discount of 10 framed two ways")
print(" tau gain frame loss frame difference (points)")
for tau in (2, 5, 10, 20, 50):
g, l = logistic(b0 + v(10) / tau), logistic(b0 + LAM * v(10) / tau)
print(f" {tau:4d} {g:10.3f} {l:10.3f} {100 * (l - g):8.1f}")
print(f" v(10) = {v(10):.2f}, lam v(10) = {LAM * v(10):.2f};"
" with lam = 1 the two frames coincide")
# gain frame: V(A) = 105.9, V(B) = 93.6; loss frame: V(C) = -438.5, V(D) = -353.1
# published share choosing the sure option (TK 1981): 72 % of 152 (A), 22 % of 155 (C)
# tau P(A) P(C) frame effect (points)
# 1 1.00 0.00 100.0
# 5 0.92 0.00 92.2
# 10 0.77 0.00 77.5
# 20 0.65 0.01 63.6
# 50 0.56 0.15 40.8
# 100 0.53 0.30 23.2
# 1000 0.50 0.48 2.4
# tau fitted to P(A) = 0.72: 13.1; it then predicts P(C) = 0.00 (observed 0.22)
#
# checkout (illustrative): base conversion 30 %, discount of 10 framed two ways
# tau gain frame loss frame difference (points)
# 2 0.950 1.000 4.9
# 5 0.661 0.929 26.7
# 10 0.478 0.703 22.5
# 20 0.385 0.502 11.6
# 50 0.333 0.376 4.3
# v(10) = 7.59, lam v(10) = 17.07; with lam = 1 the two frames coincideIn the gain frame = 105.9 > = 93.6; in the loss frame = −438.5 < = −353.1: the directions agree with the 1981 results. But shares such as “72 % against 22 %” are set entirely by : predicts 100 % against 0 %, predicts 50 % against 48 %. Setting to 13.1 reproduces the 72 % of the gain frame, but the same predicts that 0 % choose C in the loss frame, against an observed 22 %. The 1992 median parameters with one common scale cannot fit the shares in both frames at once; heterogeneity between participants, different readings of the problem and response noise all play a part.
The checkout example is more direct. Base conversion is set at 30 %, and a $10 discount is described as a gain (“save $10”) or a loss (“lose your $10 discount if you leave”). In the model the utility increments of the two frames are and , a ratio of 2.25; but the conversion gap between frames is 4.9 percentage points at , 26.7 at and 4.3 at . The gap is not monotone in : when is very small both frames push conversion close to 100 % and the gap shrinks. The “conversion gap” the framework asks for is therefore not a number that cumulative prospect theory can supply; the theory gives only a direction (if the loss frame is really coded as a loss), and the size depends on a scale parameter that must be estimated from data, and on whether customers really code “losing a discount” as a loss — which is itself what framing research has to test.
Note
Why everything here is illustrative. All conversion figures in this chapter are computed under assumed parameters; they are not predictions for any page. In practice a “loss frame” also raises a question of truthfulness: if “you will lose this offer” is not true (for example, the price does not change when the countdown ends), the wording turns from framing into misrepresentation; see Chapter 2 and the checklist in Chapter 8.
Hypothesis · not established · The marginal effect of a loss frame is larger for small offers
The author’s hypothesis (not established): for offers that are real and really expire, describing the offer as “about to be lost” rather than “available to gain” raises the purchase rate by more for small offers (below 10 % of the price) than for large ones (above 30 %), because large offers are enough to prompt purchase under either frame. What would show it wrong: in a pre-registered experiment randomising a 2 × 3 factorial of “frame × offer size”, the confidence interval for the frame-by-size interaction includes zero, or the framing effect for small offers is not larger than for large ones.
Chapter summary
The meta-analytic mean of the loss-aversion coefficient is about 1.96, but estimates vary with the task, the stakes and the model specification, and some researchers dispute that losses generally outweigh gains. The program shows a decision-maker with a true estimated at 1.43, 1.13 and 0.88 by a linear model depending on the stakes. Framing effects are reliable but small to moderate overall. Cumulative prospect theory yields values, not shares; a “conversion gap” requires a link function and a scale parameter; the 1992 medians with one common scale cannot fit both shares in the Asian disease problem, and in the checkout example the framing gap ranges from 4.3 to 26.7 percentage points. The framework’s “conversion gap” can only be illustrative.
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).