5D Library › Choice Architecture, Cumulative Prospect Theory and Computational Behavioural Economics of Pricing Interfaces · 中文
Chapter 8 Practice and Pre-Registered Testing
Note
The checklist and the experimental protocol in this chapter are for research and self-examination; they are not marketing, legal or compliance advice. The checklist cannot tell you how to set prices; it can only help you write down your assumptions before deciding and spot designs that might mislead customers. The pages, price histories and experimental data in this book are fictional; any real pricing and price indication must comply with local law, with qualified professional advice where needed.
Part 1 A four-step checklist
Chapters 1 to 6 can be condensed into a four-step checklist for examining a pricing page or checkout flow. The limits in the checklist (for example “flag if more than 10 % choose the decoy”) are examples of rules the user should write down in advance, not recommended values.
- Options. List every option and its value on each attribute. Is any option dominated by another on every listed attribute (price no lower, features no more, service no better)? If so, is it a real, purchasable option with truthful attributes? Do the number of options and the structure of their differences let customers compare (Chapters 1 and 5)?
- Reference prices. Is every struck-through or “was” price a price that was really charged? On the EU Article 6a reading, is it no higher than the lowest price in the 30 days before the reduction began? On the US Guides’ reading, was it offered to the public in good faith on a regular basis for a reasonably substantial period (Chapter 2)?
- Decoys and frames. If a dominated option is on the page, was it tested in a randomised comparison? What was the change in the target’s share, with its confidence interval? How many chose the decoy (manipulation check)? Is any “you will lose this” wording true (Chapters 3 to 5)?
- Friction and defaults. How many steps to join and to leave? Are any add-ons pre-ticked? Do time-limited offers really expire (Chapter 6)?
The program below applies the checklist to a fictitious subscription page with four plans, Lite, Plus, Pro and Plus+, two of them with struck-through prices, and a fictitious randomised test log (2,400 per arm).
Calculation
# Four-step checklist applied to a FICTITIOUS pricing page and a FICTITIOUS A/B log.
# 1 options: is any option dominated (worse or equal on every listed attribute)?
# 2 reference: is every struck-through price at most the lowest price applied in the 30
# days before the reduction began (the EU Art. 6a reading)?
# 3 decoy: in a randomised test, did adding the dominated
# option change the target's share, and how many chose the decoy (manipulation check)?
# 4 friction: steps to join versus steps to leave; is any add-on pre-ticked?
import math, statistics
PAGE = {"tiers": {"Lite": (12.0, 2, 1), "Plus": (40.0, 5, 2), "Pro": (95.0, 9, 3),
"Plus+": (48.0, 4, 2)}, # (price, features, support level)
"was": {"Plus": 59.0, "Lite": 15.0},
"history": {"Plus": [59.0] * 10 + [45.0] * 25 + [40.0] * 5, # last 40 days
"Lite": [15.0] * 32 + [12.0] * 8},
"join_steps": 3, "leave_steps": 7, "preticked": ["insurance add-on"]}
AB = {"control": (2400, 744, None), "decoy": (2400, 840, 101)} # (n, chose Plus, decoy)
def dominated(name, tiers):
p, f, s = tiers[name]
return [o for o, (q, g, t) in tiers.items() if o != name and q <= p and g >= f
and t >= s and (q, g, t) != (p, f, s)]
print("step 1 options")
for k in PAGE["tiers"]:
d = dominated(k, PAGE["tiers"])
print(f" {k:6s} {'dominated by ' + ', '.join(d) if d else 'not dominated'}")
print("step 2 reference prices (Art. 6a reading: was-price <= low of the 30 days before)")
for k, was in PAGE["was"].items():
h = PAGE["history"][k]; now = h[-1]
start = max(i for i in range(len(h)) if h[i - 1] != now) if h[0] != now else 0
low = min(h[max(0, start - 30):start])
ok = was <= low
print(f" {k:6s} was {was:5.1f}, now {now:5.1f} (since day {start - len(h)}),"
f" prior 30-day low {low:5.1f}")
print(f" announced cut {1 - now / was:.0%}:"
f"{' ok' if ok else ' FLAG, supportable cut ' + format(1 - now / low, '.0%')}")
print("step 3 decoy test (randomised, target = Plus)")
(n0, x0, _), (n1, x1, dz) = AB["control"], AB["decoy"]
p0, p1 = x0 / n0, x1 / n1; se = math.sqrt(p0 * (1 - p0) / n0 + p1 * (1 - p1) / n1)
z = statistics.NormalDist().inv_cdf(0.975)
print(f" Plus share: control {p0:.3f}, with decoy {p1:.3f};")
d = 100 * (p1 - p0); h = 100 * z * se
print(f" difference {d:+.1f} points (95 % CI {d - h:+.1f} to {d + h:+.1f})")
flag = "FLAG - dominance may not be seen" if dz / n1 > 0.10 else "ok"
print(f" decoy chosen by {dz / n1:.1%} (limit 10 %): {flag}")
print("step 4 friction and defaults")
j, l = PAGE["join_steps"], PAGE["leave_steps"]
print(f" steps to join {j}, to leave {l}: {'FLAG - asymmetric' if l > j else 'ok'}")
print(f" pre-ticked: {', '.join(PAGE['preticked']) or 'none'}"
f"{' - FLAG' if PAGE['preticked'] else ''}")
# step 1 options
# Lite not dominated
# Plus not dominated
# Pro not dominated
# Plus+ dominated by Plus
# step 2 reference prices (Art. 6a reading: was-price <= low of the 30 days before)
# Plus was 59.0, now 40.0 (since day -5), prior 30-day low 45.0
# announced cut 32%: FLAG, supportable cut 11%
# Lite was 15.0, now 12.0 (since day -8), prior 30-day low 15.0
# announced cut 20%: ok
# step 3 decoy test (randomised, target = Plus)
# Plus share: control 0.310, with decoy 0.350;
# difference +4.0 points (95 % CI +1.3 to +6.7)
# decoy chosen by 4.2% (limit 10 %): ok
# step 4 friction and defaults
# steps to join 3, to leave 7: FLAG - asymmetric
# pre-ticked: insurance add-on - FLAGStep 1 flags Plus+ as dominated by Plus (dearer, fewer features, same service): it is a decoy. The checklist does not itself judge whether a decoy is “improper” — a real, purchasable option with truthful attributes is no misrepresentation — but it requires it to be listed and its effect checked in step 3. Step 2 flags Plus: struck-through 59, now 40 (for 5 days), with a lowest price of 45 in the 30 days before the reduction, so on the Article 6a reading only 11 % of the claimed 32 % is supported; Lite went from 15 to 12 after at least 30 days at 15, so its claimed 20 % is supported. Step 3: with the decoy, Plus’s share rises from 0.310 to 0.350, a difference of +4.0 points (95 % confidence interval +1.3 to +6.7); 4.2 % chose the decoy, within the pre-set 10 %, so the test is interpretable. Step 4 flags two items: leaving takes 7 steps against 3 to join, and an insurance add-on is pre-ticked. The lesson of the example is that a page can “work” in step 3 while having problems in steps 2 and 4; the checklist keeps them apart.
Part 2 A pre-registered online randomised experiment
Question and design choices
On a three-tier pricing page for a fictitious product, run by independent researchers, does adding an option asymmetrically dominated by the target tier raise the share of people choosing the target tier, by at least the smallest effect of interest?
The author’s framework proposes comparing a “no decoy” page with one that “adds an asymmetrically dominated decoy”. The book keeps these arms and makes four design choices. First, a fictitious product: brand, plan names and prices on the page are all fictitious (three tiers at US$12, 39 and 89, and a decoy at 45 slightly worse than the 39 tier), using neither the author’s products nor the prices 9.99, 49.90 and 149.00 that match 5D Library’s own, so that the result cannot be used to endorse the author’s pricing. Second, attributes shown side by side in numbers: the evidence of Chapter 5 shows that the attraction effect is most likely when dominance is visible at a glance with numerical attributes; this experiment tests “does the effect reach a size worth acting on in the most favourable presentation”, not “is the decoy effect universal”. Third, a single primary outcome: the share choosing the target (US$39) tier, a binary outcome; shares of the decoy, basic and premium tiers and other measures are secondary. Fourth, run by independent researchers: the protocol, page files and analysis code are registered publicly before recruitment, conduct and analysis are by researchers unconnected with the author’s businesses, and nobody is recruited from the author’s community, students or customers.
Protocol
| Element | Specification |
|---|---|
| Participants | Adults recruited by the independent team through a general online survey panel; excluding employees, clients, students and community members of the author’s businesses (panel screening question plus self-declaration); those failing the pre-registered attention check are excluded |
| Randomisation and stratification | 1:1 allocation to A (three tiers) or B (three tiers + decoy), stratified by device (phone / computer); the random sequence is generated by the platform and hidden from researchers before allocation |
| Blinding | Participants do not know another page exists; analysts see only “X / Y” labels until the results are locked; the author has no access to raw data or labels |
| Incentive compatibility | Each person has a budget of 100 experimental credits, page prices are shown in credits (1 credit = US$1, so US$12/39/45/89 are 12/39/45/89 credits), and each chooses one tier or none; one in ten, drawn at random, is paid: a real third-party gift of equivalent value for the chosen plan (chosen by the research team and unconnected with the author) plus the remaining budget in cash (100 minus the chosen price, at US$1 per credit; those choosing none receive US$100). With a single choice paid with a random draw, this is incentive compatible under monotonicity (Azrieli et al. 2018) |
| Primary outcome | Difference in the share choosing the target tier (B − A, percentage points); 95 % confidence interval by Newcombe’s Wilson-score method (Newcombe 1998) |
| Smallest effect of interest | 5 percentage points (rationale below) |
| Secondary outcomes | Share choosing the decoy (also the manipulation check); shares of the basic and premium tiers; share choosing none; time to choose; confidence in the chosen plan; the primary contrast by device; all descriptive or exploratory |
| Non-interpretability gates | Decoy chosen by more than 10 % in B; exclusion rates for the attention check differing by more than 3 points between arms; arm sizes imbalanced (chi-square ); the page differing from the registered files |
Smallest effect of interest and sample size
The smallest effect of interest needs a rationale. The evidence of Chapter 5 gives two quantities: in laboratory studies with numerical attributes, adding a decoy moves the target’s share (over the whole sample) by about 14 to 24 percentage points (Frederick et al. 2014); in field data from an online diamond retailer, only 11 % to 25 % of customers notice dominance (Wu and Cosguner 2020). Their product puts the average effect on a page at roughly 1.5 to 6.0 points. This is only a rough order of magnitude: the laboratory figures are already whole-sample shifts (including people who did not notice), so multiplying by a detection rate discounts twice and the range may be too low; a detection rate from a diamond market need not carry over to a subscription page either. This protocol takes 5 points, in the upper part of that range, for two reasons: the experiment uses the side-by-side numerical presentation most favourable to noticing, so the detection rate should be at the high end; and a change smaller than 5 points would not pay for the trust and maintenance costs of carrying an “unattractive” option on the page. The trade-off means that if the true effect is only about 2 points, the experiment will most likely conclude “no difference worth acting on” — one of the conclusions it is meant to test.
Calculation
# Sample size for the Chapter 8 pre-registered online experiment. Arms: A = three tiers
# of a fictitious subscription (12 / 39 / 89 dollars, NOT the author's products); B = the
# same page plus an asymmetrically dominated decoy next to the middle (target) tier.
# Single primary outcome: chose the target tier (binary). Two-sided alpha 0.05,
# power 90 %.
# n per arm = (z_a sqrt(2 pbar qbar) + z_b sqrt(p1 q1 + p2 q2))^2 / (p2 - p1)^2.
import math, random, statistics
Z = statistics.NormalDist().inv_cdf
ZA, ZB = Z(0.975), Z(0.90)
def n_arm(p1, p2):
pb = (p1 + p2) / 2
num = (ZA * math.sqrt(2 * pb * (1 - pb))
+ ZB * math.sqrt(p1 * (1 - p1) + p2 * (1 - p2))) ** 2
return math.ceil(num / (p2 - p1) ** 2)
print("rough order of magnitude for the SESOI: lab shift x field detection rate")
print(" (whole-sample shifts in numeric lab studies, Frederick et al. 2014: +14, +16,")
print(" +24 points; detection 11-25 % in Wu & Cosguner 2020's diamond data)")
for d in (0.11, 0.25):
print(f" detection {d:.0%}: " + ", ".join(f"{d * e:.1f}" for e in (14, 16, 24))
+ " points")
P1, D = 0.30, 0.05
n = n_arm(P1, P1 + D)
plan = math.ceil(n / 0.85 / 100) * 100
print(f"base share of the target tier {P1:.0%}, SESOI {100 * D:.0f} points:"
f" {n:,} per arm analysable")
print(f" allowing 15 % exclusions (attention checks, incomplete sessions):")
print(f" recruit {plan:,} per arm, {2 * plan:,} in all")
rng = random.Random(47); hits = alpha = 0; NSIM = 10000
for _ in range(NSIM): # simulation check of the formula
for p2, null in ((P1 + D, False), (P1, True)):
a = sum(rng.random() < P1 for _ in range(n)) / n # Bernoulli sums:
b = sum(rng.random() < p2 for _ in range(n)) / n # portable to Python 3.8+
se = math.sqrt(a * (1 - a) / n + b * (1 - b) / n)
sig = abs(b - a) / se > ZA
alpha += sig and null; hits += sig and not null
se_p, se_a = math.sqrt(0.9 * 0.1 / NSIM), math.sqrt(0.05 * 0.95 / NSIM)
q1, q2, pb = P1 * (1 - P1), (P1 + D) * (1 - P1 - D), P1 + D / 2
exact = statistics.NormalDist().cdf((D * math.sqrt(n) - ZA * math.sqrt(2 * pb * (1 - pb)))
/ math.sqrt(q1 + q2))
print(f" power at n = {n:,} by the normal approximation: {exact:.3f}")
print(f" {NSIM:,} simulated trials: power {hits / NSIM:.3f} (MC s.e. {se_p:.3f}),")
print(f" type I error {alpha / NSIM:.3f} (MC s.e. {se_a:.4f})")
print("sensitivity: analysable n per arm SESOI 3 pts 5 pts 8 pts")
print(" (pre-set rule: if the pilot base share is not 30 %, recompute n for SESOI 5")
print(" at the pilot share and use the larger of that and 1,842)")
for p1 in (0.20, 0.30, 0.40):
print(f" base share {p1:.0%} " + "".join(
f"{n_arm(p1, p1 + d):8,d}" for d in (0.03, 0.05, 0.08)))
# rough order of magnitude for the SESOI: lab shift x field detection rate
# (whole-sample shifts in numeric lab studies, Frederick et al. 2014: +14, +16,
# +24 points; detection 11-25 % in Wu & Cosguner 2020's diamond data)
# detection 11%: 1.5, 1.8, 2.6 points
# detection 25%: 3.5, 4.0, 6.0 points
# base share of the target tier 30%, SESOI 5 points: 1,842 per arm analysable
# allowing 15 % exclusions (attention checks, incomplete sessions):
# recruit 2,200 per arm, 4,400 in all
# power at n = 1,842 by the normal approximation: 0.900
# 10,000 simulated trials: power 0.895 (MC s.e. 0.003),
# type I error 0.054 (MC s.e. 0.0022)
# sensitivity: analysable n per arm SESOI 3 pts 5 pts 8 pts
# (pre-set rule: if the pilot base share is not 30 %, recompute n for SESOI 5
# at the pilot share and use the larger of that and 1,842)
# base share 20% 3,939 1,464 597
# base share 30% 5,037 1,842 735
# base share 40% 5,667 2,053 807The baseline share of the target tier is assumed to be 30 % (to be checked by the independent team in a small pilot not counted in the analysis). With two-sided , 90 % power and a smallest effect of interest of 5 points, 1,842 analysable participants per arm are needed; allowing for 15 % exclusions (attention checks and incomplete sessions), 2,200 are recruited per arm, 4,400 in all. At this sample size the normal approximation gives a power of 0.900; in 10,000 simulated trials (each arm drawn as a sum of Bernoulli trials, runnable on Python 3.8 and later) the power is 0.895 (Monte Carlo standard error 0.003) and the type I error rate 0.054 (standard error 0.0022), both within about two Monte Carlo standard errors of the nominal 0.90 and 0.05; the formal analysis uses the Newcombe interval. The sensitivity table shows that a smallest effect of interest of 3 points would need 5,037 per arm, and a baseline of 40 % would need 2,053 for 5 points. Pre-set rule: if the pilot finds a baseline share different from 30 %, the analysable sample per arm is recomputed with the same formula at the pilot share (smallest effect of interest still 5 points), the larger of that and 1,842 is used, and recruitment is scaled for 15 % exclusions; this adjustment is made and registered before main recruitment begins and is not changed afterwards.
By the result in Book 33, “90 % power” means that if the true effect equals the smallest effect of interest, a conclusion of the kind “the decoy raised the target share” (lower confidence limit above 0) is reached with probability about 90 %; in that case the strongest conclusion, “raised by at least the smallest effect of interest”, has only about a 2.5 % chance.
Guarding against data snooping
There is one pre-specified page design and one primary outcome; the protocol, page files (with screenshots and hashes), randomisation procedure and analysis code are registered publicly before recruitment (Nosek et al. 2018); there is no interim analysis and no early stopping; secondary and exploratory results are labelled as such and cannot replace the primary conclusion. Book 33 discussed the statistical rationale for these practices.
Decision rule
Calculation
# Decision rule, fixed before the independent team unblinds the data. Primary outcome:
# share choosing the target tier, B (with decoy) minus A (three tiers), in points, with a
# 95 % Newcombe (Wilson score) interval. SESOI 5 points. Gates, checked first: decoy
# chosen by more than 10 % in B (dominance not perceived as designed); attention-check
# exclusions differing between arms by more than 3 points; randomisation imbalance
# (chi-square p < 0.001 on arm sizes); any deviation from the registered page files.
import math
SESOI, Z = 5.0, 1.959963984540054
def wilson(x, n):
p = x / n; c = Z * Z / n
mid = (p + c / 2) / (1 + c)
half = Z * math.sqrt(p * (1 - p) / n + c / (4 * n)) / (1 + c)
return mid - half, mid + half
def newcombe(x1, n1, x2, n2):
p1, p2 = x1 / n1, x2 / n2; l1, u1 = wilson(x1, n1); l2, u2 = wilson(x2, n2)
d = p2 - p1
return (100 * d, 100 * (d - math.sqrt((p2 - l2) ** 2 + (u1 - p1) ** 2)),
100 * (d + math.sqrt((u2 - p2) ** 2 + (p1 - l1) ** 2)))
def verdict(xa, na, xb, nb, decoy=0.02, excl_gap=0.0, deviation=False):
d, lo, hi = newcombe(xa, na, xb, nb)
chi = (na - nb) ** 2 / (na + nb) # 1 df, equal allocation
for bad, why in ((decoy > 0.10, f"decoy chosen by {decoy:.0%} (limit 10 %)"),
(abs(excl_gap) > 3, f"exclusions differ by {abs(excl_gap):.1f}"
" points"),
(chi > 10.83, "arm sizes imbalanced (p < 0.001)"),
(deviation, "page differs from registered files")):
if bad:
return f"{d:+5.1f}: NOT INTERPRETABLE - {why}"
if lo >= SESOI: v = "decoy raises the target share by at least the SESOI"
elif lo > 0: v = "decoy raises the target share; SESOI not shown"
elif hi < 0: v = "decoy LOWERS the target share (repulsion)"
elif lo >= -SESOI and hi <= SESOI: v = "no difference worth acting on"
else: v = "unclear: interval too wide"
return f"{d:+5.1f} ({lo:+5.1f} to {hi:+5.1f}): {v}"
for case in ((553, 1842, 700, 1842), (553, 1842, 645, 1842), (553, 1842, 560, 1842),
(553, 1842, 480, 1842), (60, 200, 70, 200), (553, 1842, 700, 1842, 0.14),
(553, 1842, 700, 1842, 0.02, 4.1), (553, 1842, 700, 1600),
(553, 1842, 790, 1842)):
print(verdict(*case))
print("secondary outcomes (decoy share, premium, basic and 'none' shares, time to")
print(" choose, stated confidence, the same contrast by device type) carry intervals")
print(" only; they never change the verdict")
# +8.0 ( +4.9 to +11.0): decoy raises the target share; SESOI not shown
# +5.0 ( +2.0 to +8.0): decoy raises the target share; SESOI not shown
# +0.4 ( -2.6 to +3.3): no difference worth acting on
# -4.0 ( -6.9 to -1.1): decoy LOWERS the target share (repulsion)
# +5.0 ( -4.2 to +14.0): unclear: interval too wide
# +8.0: NOT INTERPRETABLE - decoy chosen by 14% (limit 10 %)
# +8.0: NOT INTERPRETABLE - exclusions differ by 4.1 points
# +13.7: NOT INTERPRETABLE - arm sizes imbalanced (p < 0.001)
# +12.9 ( +9.8 to +15.9): decoy raises the target share by at least the SESOI
# secondary outcomes (decoy share, premium, basic and 'none' shares, time to
# choose, stated confidence, the same contrast by device type) carry intervals
# only; they never change the verdictThe example results are invented. If any gate fails (decoy chosen by more than 10 %, exclusion rates differing by more than 3 points, imbalanced arms, a page differing from the registered files), the report says “not interpretable” whatever the primary outcome. Only when the lower confidence limit is at least 5 points does the report say “the decoy raised the target share by at least the smallest effect of interest”; when the lower limit is above 0 but below 5, it says “the decoy raised the target share, but has not been shown to reach the smallest effect of interest” (the first example, with an estimate of +8.0 but a lower limit of 4.9, falls here); when the upper limit is below 0, “the decoy lowered the target share (repulsion)”; when the interval lies within ±5, “no difference worth acting on”; otherwise, “unclear”.
What this experiment cannot show
First, it tests one presentation (side-by-side numerical attributes), one price structure and one decoy position, and cannot be generalised to pages with pictures, experiential attributes or attributes that need calculation to compare; Chapter 5’s evidence suggests the effect may vanish or reverse there. Second, participants choose with experimental credits and third-party gifts, not by buying a long-term subscription with their own money in a real shop; incentive compatibility only gives them a reason to choose by preference, not identical behaviour to real purchases. Third, it does not test Chapter 2’s struck-through prices, Chapter 4’s loss frames or Chapter 6’s form friction; these need their own experiments. Fourth, it does not test the pricing of 5D Library or any real product; even a positive result would not mean that the prices 9.99, 49.90 and 149.00 “work”. Fifth, it measures one choice, not cancellations, satisfaction or long-term trust; a decoy that raises the immediate share might damage trust in the long run, which needs a different design. Sixth, online-panel participants do not represent the customers of any particular market. If the experiment concludes “no difference worth acting on” or “repulsion”, that is its conclusion.
Ethics and conflicts of interest
The experiment involves human participants and needs approval from the conducting institution’s ethics committee; participants take part after informed consent and may withdraw at any time; the brand and plans on the page are fictitious, and participants are told so, and told the purpose of the study, at the end (debriefing); the gifts paid are chosen by the research team and unconnected with the author or his businesses. Only the data the experiment needs are collected, with no personally identifying information; de-identified data are published. Whatever the result, positive or negative, the results, code and de-identified data will be published. To limit the influence of interests: the author takes no part in recruitment, conduct or analysis and has no access to raw data or arm labels; the experiment uses neither the author’s products nor prices matching them; the protocol and code are registered publicly before recruitment and the results will be published whatever they are.
Disclosure
The author uses the works of 5D Library, including this book, as the methodological basis of his community courses, the 5D Ascension Cards (Dimensional Shift Card Decks) and the Yafu AI knowledge base, and therefore has a commercial interest in the subject of this book. The book promotes no course, card deck or service; the Chapter 8 trial is designed to be run and analysed by researchers independent of the author’s business and does not recruit from the author’s community, students or customers; all the code and the study protocol are published openly for anyone to check. The author founded and runs Asia Influencer X and Yafu Global, which provide influencer marketing, live-commerce and digital marketing services to brands; that work involves the pricing pages and choice design discussed in this book, so he has a direct commercial interest in its subject. The three-tier example prices in the book (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, and the author has a direct interest in that price structure; the book’s analysis does not validate those prices. The book uses no data from the author’s businesses or their clients as evidence and does not promote their services.
Chapter summary
The four-step checklist examines a pricing page through options, reference prices, decoys and frames, and friction and defaults; the example shows that a page can “work” in a decoy test while carrying an unsupported struck-through price and an asymmetric exit route. The pre-registered online randomised experiment, run by independent researchers on a three-tier page for a fictitious product, compares the share choosing the target tier with and without an asymmetrically dominated decoy; the single primary outcome is their difference, the smallest effect of interest 5 percentage points, with 1,842 analysable and 2,200 recruited per arm for 90 % power; the decision rule is written as code, and results are not interpretable if too many choose the decoy or exclusions are imbalanced; they will be published whatever they are. It tests one design in one presentation, does not test the author’s pricing and cannot be carried over directly to real long-term purchases.
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).