5D Library › Choice Architecture, Cumulative Prospect Theory and Computational Behavioural Economics of Pricing Interfaces · 中文

Chapter 1 How Many Options: Choice Overload and the Size of Nudge Effects

The original numbers of the jam study

Evidence

Study 1 of Iyengar and Lepper (2000) was run in an upscale supermarket in Menlo Park, California: a tasting booth displayed either 24 or 6 jams in rotation, passing shoppers could taste them and received a coupon for $1 off any jam. Of the 242 customers who passed the 24-jam display, 145 (60 %) stopped; of the 260 who passed the 6-jam display, 104 (40 %) stopped. Of those who stopped, 4 (about 3 %) in the 24-jam condition later bought with the coupon, against 31 (about 30 %) in the 6-jam condition (Iyengar and Lepper 2000). In Study 3 of the same paper, 48 % of participants who chose from 6 chocolates took chocolates rather than $5 in cash as payment, against 12 % of those who chose from 30.

The program below takes these published counts, computes purchase rates per stopper and per passer-by, and runs Fisher’s exact test.

Calculation

# Part A: the published counts of Iyengar & Lepper (2000), Study 1 (jam tasting booth),
# re-analysed per stopper and per passer-by, with Fisher's exact test.
# Part B: THE AUTHOR'S MODEL (illustrative, not evidence). A shopper sees n options whose
# match values u_i are iid N(0, 1). With preference uncertainty she observes u_i + e_i,
# e_i ~ N(0, s^2), so the best posterior mean is r * M_n with r = 1 / sqrt(1 + s^2) and
# M_n the maximum of n standard normals. Reviewing the set costs c per extra option. She
# buys if r * M_n - c * (n - 1) > u0, otherwise she defers. P(buy) is exact:
# P(M_n > t) = 1 - Phi(t) ** n with t = (u0 + c (n - 1)) / r.
import math, statistics

ND = statistics.NormalDist()

def fisher_two_sided(a, n1, b, n2):
    k, n = a + b, n1 + n2
    p = lambda x: math.comb(n1, x) * math.comb(n2, k - x) / math.comb(n, k)
    p0 = p(a)
    xs = range(max(0, k - n2), min(k, n1) + 1)
    return sum(p(x) for x in xs if p(x) <= p0 * 1.0000001)

def p_buy(n, c, s, u0=1.0):
    r = 1 / math.sqrt(1 + s * s)
    return 1 - ND.cdf((u0 + c * (n - 1)) / r) ** n

print("Iyengar & Lepper 2000, Study 1 (published counts)")
for lab, passed, stop, buy in (("24 jams", 242, 145, 4), ("6 jams", 260, 104, 31)):
    print(f"   {lab}: stopped {stop}/{passed} = {stop / passed:.0%}; bought {buy}/{stop}"
          f" = {buy / stop:.1%} of stoppers, {buy / passed:.1%} of passers-by")
print("   Fisher exact p, buyers among stoppers:",
      f"{fisher_two_sided(4, 145, 31, 104):.1e};"
      f" among passers-by: {fisher_two_sided(4, 242, 31, 260):.1e}")
print(f"   ratio of purchase rates per passer-by: {(31 / 260) / (4 / 242):.1f}")
print()
print("author's model: P(buy) for 6 and 24 options, best n in 1..60")
print("   cost c  noise s    n=6    n=24   24 minus 6   best n  P(best)")
effects = []
for c in (0.0, 0.005, 0.02, 0.05):
    for s in (0.0, 1.0, 2.0):
        p6, p24 = p_buy(6, c, s), p_buy(24, c, s)
        best = max(range(1, 61), key=lambda n: p_buy(n, c, s))
        effects.append(p24 - p6)
        print(f"   {c:6.3f}  {s:7.1f}  {p6:6.3f}  {p24:6.3f}  {p24 - p6:+10.3f}"
              f"  {best:7d}  {p_buy(best, c, s):6.3f}")
pos = sum(e > 0 for e in effects)
print(f"12 settings: 24 beats 6 in {pos}, 6 beats 24 in {12 - pos};"
      f" mean difference {statistics.fmean(effects):+.3f},"
      f" sd {statistics.stdev(effects):.3f}")
# Iyengar & Lepper 2000, Study 1 (published counts)
#    24 jams: stopped 145/242 = 60%; bought 4/145 = 2.8% of stoppers, 1.7% of passers-by
#    6 jams: stopped 104/260 = 40%; bought 31/104 = 29.8% of stoppers, 11.9% of passers-by
#    Fisher exact p, buyers among stoppers: 1.8e-09; among passers-by: 4.4e-06
#    ratio of purchase rates per passer-by: 7.2
#
# author's model: P(buy) for 6 and 24 options, best n in 1..60
#    cost c  noise s    n=6    n=24   24 minus 6   best n  P(best)
#     0.000      0.0   0.645   0.984      +0.339       60   1.000
#     0.000      1.0   0.388   0.860      +0.472       60   0.993
#     0.000      2.0   0.074   0.264      +0.190       60   0.535
#     0.005      0.0   0.630   0.967      +0.337       60   0.998
#     0.005      1.0   0.368   0.758      +0.390       60   0.871
#     0.005      2.0   0.064   0.141      +0.077       31   0.146
#     0.020      0.0   0.583   0.834      +0.251       25   0.835
#     0.020      1.0   0.310   0.376      +0.067       16   0.415
#     0.020      2.0   0.041   0.013      -0.028        8   0.042
#     0.050      0.0   0.488   0.317      -0.171       10   0.534
#     0.050      1.0   0.210   0.028      -0.182        6   0.210
#     0.050      2.0   0.015   0.000      -0.015        3   0.021
# 12 settings: 24 beats 6 in 8, 6 beats 24 in 4; mean difference +0.144, sd 0.219

The first part of the program makes three points. First, the often-quoted “3 % against 30 %” is the purchase share among those who stopped (2.8 % against 29.8 %); per passer-by it is 1.7 % against 11.9 %, a ratio of about 7.2. Both favour the 6-jam display, but only the second answers the commercial question “of the people who walk past this booth, how many buy”. Second, the difference is highly significant (exact p-values of 1.8×10−91.8\times10^{-9} and 4.4×10−64.4\times10^{-6}), but it comes from one field study; the rotation times, the coupon’s validity and the mix of shoppers were not replicated across stores. Third, the 24-jam display drew more people in, so “more options” worked in opposite directions on “attracting attention” and “closing the sale”; folding both into one “conversion rate” hides this.

Two meta-analyses

Evidence

Scheibehenne, Greifeneder and Todd (2010) pooled 63 conditions from 50 published and unpublished experiments (5,036 participants in all) and found a mean choice-overload effect of “virtually zero”, with considerable variation between studies; they identified several possible preconditions but no sufficient conditions under which the effect reliably appears (Scheibehenne et al. 2010). Chernev, Böckenholt and Goodman (2015) pooled 99 observations (7,202 participants) and identified four moderators: choice-set complexity, decision-task difficulty, preference uncertainty and an effort-minimising decision goal; the stronger each, the stronger the overload. With these moderators taken into account, the overall effect of assortment size on choice overload was significant, which the authors present as differing from the earlier meta-analysis (Chernev et al. 2015).

The two meta-analyses do not contradict each other. The first asks “is there an effect on average?” and answers “close to zero”; the second asks “when does the effect appear?” and answers with four moderators. Together they say that choice overload is not a general law and not non-existent; it is a conditional effect. Both “more options, more conversions” and “too many options and nobody buys” should therefore be rewritten as conditional statements.

The author’s model: two faces of the number of options

To show how conditions set the direction, here is a simple author’s model. It is not fitted to any study; it only illustrates a mechanism.

A shopper faces nn options whose match uiu_i with her needs is independent standard normal. She cannot see the true match, only a noisy signal ui+eiu_i + e_i, ei∼N(0,s2)e_i \sim N(0, s^2); her posterior estimate of the best option is therefore rMnr M_n, where MnM_n is the maximum of nn standard normals and r=1/1+s2r = 1/\sqrt{1 + s^2} reflects preference uncertainty (the larger ss, the harder it is to tell good from bad). Comparing options one by one costs cc per extra option. She buys if rMn−c(n−1)>u0r M_n – c(n-1) > u_0 and otherwise defers (u0=1u_0 = 1 is the value of not buying). The purchase probability is exact:

P(buy)=1−Φ(u0+c(n−1)r)n.P(\text{buy}) = 1 – \Phi\!\left(\frac{u_0 + c(n-1)}{r}\right)^{n}.

More options act in two ways: MnM_n grows with nn (a better match becomes more likely) and the comparison cost c(n−1)c(n-1) grows linearly. The former grows ever more slowly (roughly like 2ln⁡n\sqrt{2\ln n}), the latter does not slow down, so as long as c>0c > 0 the purchase probability eventually falls with nn.

The table in the second part of the program lists 12 settings. With zero comparison cost, 24 options always beat 6 (with preference uncertainty s=1s = 1, 0.860 against 0.388; here “best nn = 60” is only the search cap of 60, since the purchase probability rises monotonically in nn); with a cost of 0.05, 24 always lose (with s=1s = 1, 0.028 against 0.210), and with s=1s = 1 the best number of options falls to 6 (10 for s=0s = 0, 3 for s=2s = 2). In the intermediate settings, the greater the preference uncertainty, the smaller the best number of options (with cost 0.02, as ss goes from 0 to 2 the best number falls from 25 to 8). Across the 12 settings, 24 options do better in 8 and worse in 4; the mean difference is +0.144 with a standard deviation of 0.219, so the average hides cases of opposite sign. This matches the picture from the meta-analyses: comparison cost corresponds to “task difficulty” and “complexity”, noise to “preference uncertainty”.

Note

What this model cannot show. It assumes independent matches, one-by-one comparison and linear costs, and it leaves out the effect of attracting people to stop (more people stopped at the 24-jam display). It shows only that in a reasonable mechanism the direction of the effect is set by comparison cost and preference uncertainty, so any single “best number of options” has to be tested in context.

Left: purchase probability against the number of options in the author’s model (preference uncertainty s = 1, different comparison costs; dotted lines at 6 and 24). Right: when only significant results are published, the mean published estimate against the sample size per arm (base rate 20 %, three true effects; dotted lines show the true effects). Illustrative.

How much of a nudge survives at scale

Choice overload is just one phenomenon of “choice architecture”. Thaler and Sunstein define choice architecture as the way options are presented, and call a “nudge” any design that predictably changes behaviour without forbidding options or significantly changing economic incentives (Thaler and Sunstein 2008). Nudge effects are often sizeable in laboratory and small field studies and much smaller when rolled out at scale.

Evidence

DellaVigna and Linos (2022) assembled all 126 randomised controlled trials of two US nudge units (covering more than 23 million people): their average effect was a take-up increase of 1.4 percentage points (8.0 % over the control group), against 8.7 percentage points (33.4 %) for nudge trials published in academic journals. They attribute most of the gap to selective publication in the academic literature, while also discussing differences in interventions and populations between the two sets (DellaVigna and Linos 2022). Mertens and co-authors’ 2022 meta-analysis reported a medium average effect (Cohen’s dd about 0.43) (Mertens et al. 2022); Maier and co-authors reanalysed the same data with Bayesian model averaging that corrects for publication bias and obtained an adjusted overall effect of 0.04 (95 % interval 0.00 to 0.14), concluding that “no evidence remains” that nudges are effective overall, without ruling out that some categories are (Maier et al. 2022). Defaults are a category with relatively strong evidence: Jachimowicz and co-authors pooled 58 studies (73,675 participants) and found a mean default effect of d=0.68d = 0.68 (95 % interval 0.53 to 0.83), stronger in consumer domains and weaker in environmental ones (Jachimowicz et al. 2019).

Why? The program below gives a purely arithmetical illustration: suppose every trial has the same true effect of 1.4 percentage points (base rate 20 %), nn people per arm, and only results with z>1.96z > 1.96 are published. The expectation of a truncated normal gives the mean published estimate

E[d̂∣d̂>1.96se]=d+seϕ(a)1−Φ(a),a=1.96−d/se.E[\hat d \mid \hat d > 1.96\,\mathrm{se}] = d + \mathrm{se}\,\frac{\phi(a)}{1 – \Phi(a)}, \qquad a = 1.96 – d/\mathrm{se}.

Calculation

# Why published nudge effects shrink at scale: a selection-on-significance calculation.
# DellaVigna & Linos (2022): trials in academic journals average 8.7 percentage points,
# all trials of two US nudge units 1.4 points. Here, purely as arithmetic, suppose every
# trial has the same true effect d (points) on a 20 % base rate, n per arm, and only
# results with z > 1.96 are published. The mean published estimate is
#   E[est | est > 1.96 se] = d + se * phi(a) / (1 - Phi(a)),  a = 1.96 - d / se,
# se = sqrt(2 p (1 - p) / n) in points. Exact normal formulas; nothing is simulated.
import math, statistics

ND = statistics.NormalDist()

def published(d, n, p=0.20):
    se = 100 * math.sqrt(2 * p * (1 - p) / n)
    a = 1.96 - d / se
    return d + se * ND.pdf(a) / (1 - ND.cdf(a)), 1 - ND.cdf(a), se

print("true effect d = 1.4 points on a 20 % base rate; only z > 1.96 published")
print("   n per arm    s.e.   power   mean published   exaggeration")
for n in (100, 250, 500, 1000, 2500, 10000, 100000):
    m, pw, se = published(1.4, n)
    print(f"   {n:9,d}  {se:6.2f}  {pw:6.3f}  {m:15.2f}  {m / 1.4:12.1f}x")
lo, hi = 10, 100000
for _ in range(60):
    mid = math.sqrt(lo * hi)
    lo, hi = (mid, hi) if published(1.4, mid)[0] > 8.7 else (lo, mid)
print(f"   a published mean of 8.7 arises at about n = {lo:,.0f} per arm")
print("if instead the true effect is 8.7 points (no filter needed):")
for n in (250, 1000):
    m, pw, se = published(8.7, n)
    print(f"   n = {n:5,d}: power {pw:.3f}, mean published {m:.2f}")
# true effect d = 1.4 points on a 20 % base rate; only z > 1.96 published
#    n per arm    s.e.   power   mean published   exaggeration
#          100    5.66   0.043            13.40           9.6x
#          250    3.58   0.058             8.55           6.1x
#          500    2.53   0.080             6.10           4.4x
#        1,000    1.79   0.120             4.39           3.1x
#        2,500    1.13   0.235             2.88           2.1x
#       10,000    0.57   0.697             1.68           1.2x
#      100,000    0.18   1.000             1.40           1.0x
#    a published mean of 8.7 arises at about n = 241 per arm
# if instead the true effect is 8.7 points (no filter needed):
#    n =   250: power 0.681, mean published 10.57
#    n = 1,000: power 0.998, mean published 8.71

With 100 per arm, power is only 4.3 % and the mean published estimate is 13.40 points, 9.6 times the true effect; at about 241 per arm the mean published value is exactly 8.7; at 10,000 per arm the exaggeration falls to 1.2. This does not say that DellaVigna and Linos’s data “were” generated this way — academic and nudge-unit trials really do differ in intensity, population and outcome — but it shows that with small samples and publication of significant results only, a true effect of 1.4 points is enough to appear as 8.7 in the literature. Book 33 treated the same mechanism (the winner’s curse); here it serves as a reminder: on seeing “this page change lifted conversion 30 %”, first ask how large the sample was, whether it was registered in advance and whether the failed versions were published too.

Hypothesis · not established · Comparison cost sets the direction of the option-count effect

The author’s hypothesis (not established): on a pricing page, the direction in which more plans affect purchases depends on comparison cost: when plans differ along one directly comparable dimension (for example only storage), more options raise the purchase rate; when differences are spread across several hard-to-compare dimensions, more options lower it. What would show it wrong: in a pre-registered factorial experiment randomising “one-dimensional” and “multi-dimensional” pages and comparing 3 with 6 plans in each, the confidence interval for the interaction between number of options and difference structure includes zero or has the opposite sign.

From evidence to pages: three things to write down when testing the number of options

Applying the discussion above to a real page requires writing down at least three things in advance. First, the denominator: the jam study shows that “purchase rate among those who stopped” and “purchase rate among those who passed” can give very different ratios; the primary outcome of a page test should take as its denominator “visitors randomly assigned to the page”, not “people who opened the plan details”, or the effects on attracting attention and on closing the sale get mixed. Second, the number of outcomes: the number of options may change the purchase rate, the average price of the plan chosen and the refund rate at once; only one can be primary, the others secondary (as in Chapter 8). Third, the structure of the comparison: the author’s model shows that the direction of the effect depends on comparison cost, so a “3 plans against 6” result cannot be carried over to a different set of 6 with a different structure of differences; one test answers only for the pages it tested. Book 6 wrote the conversion funnel as an absorbing Markov chain and discussed the elasticity of each stage; here we add only that the number of options often changes not one stage but two transition probabilities at once — “moving to the next stage” and “staying at this stage” — and the final conversion rate alone cannot tell them apart.

Chapter summary

The “3 % against 30 %” of the jam study is the purchase share among those who stopped; per passer-by it is 1.7 % against 11.9 %, and the 24-jam display drew more people in. Meta-analysis shows a mean choice-overload effect near zero with large heterogeneity, and identifies four moderators: complexity, task difficulty, preference uncertainty and an effort-minimising goal. The author’s model shows the direction of the option-count effect set by comparison cost and preference uncertainty, favouring more options in 8 of 12 settings and fewer in 4. Large nudge-unit trials average 1.4 percentage points against 8.7 in academic journals; with only significant results published, a true effect of 1.4 points appears on average as 8.7 in studies of about 240 per arm. So neither “more options are better” nor “choice overload is universal” holds, and any large effect reported from a single experiment calls first for the question of what filter it passed through.

References

Chernev, Alexander, Ulf Böckenholt, and Joseph Goodman. 2015. “Choice Overload: A Conceptual Review and Meta-Analysis.” Journal of Consumer Psychology 25 (2): 333–58.
DellaVigna, Stefano, and Elizabeth Linos. 2022. “RCTs to Scale: Comprehensive Evidence from Two Nudge Units.” Econometrica 90 (1): 81–116. https://doi.org/10.3982/ECTA18709.
Iyengar, Sheena S., and Mark R. Lepper. 2000. “When Choice Is Demotivating: Can One Desire Too Much of a Good Thing?” Journal of Personality and Social Psychology 79 (6): 995–1006. https://doi.org/10.1037/0022-3514.79.6.995.
Jachimowicz, Jon M., Shannon Duncan, Elke U. Weber, and Eric J. Johnson. 2019. “When and Why Defaults Influence Decisions: A Meta-Analysis of Default Effects.” Behavioural Public Policy 3 (2): 159–86.
Maier, Maximilian, František Bartoš, T. D. Stanley, David R. Shanks, Adam J. L. Harris, and Eric-Jan Wagenmakers. 2022. “No Evidence for Nudging after Adjusting for Publication Bias.” Proceedings of the National Academy of Sciences 119 (31): e2200300119. https://doi.org/10.1073/pnas.2200300119.
Mertens, Stephanie, Mario Herberz, Ulf J. J. Hahnel, and Tobias Brosch. 2022. “The Effectiveness of Nudging: A Meta-Analysis of Choice Architecture Interventions across Behavioral Domains.” Proceedings of the National Academy of Sciences 119 (1): e2107346118. https://doi.org/10.1073/pnas.2107346118.
Scheibehenne, Benjamin, Rainer Greifeneder, and Peter M. Todd. 2010. “Can There Ever Be Too Many Options? A Meta-Analytic Review of Choice Overload.” Journal of Consumer Research 37 (3): 409–25. https://doi.org/10.1086/651235.
Thaler, Richard H., and Cass R. Sunstein. 2008. Nudge: Improving Decisions about Health, Wealth, and Happiness. Yale University Press.

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).