Does counting change what counts? Quantification fixation biases decision-making

Quantification increasingly permeates everyday judgments, from newborn Apgar scores to product ratings and costs. The paper asks whether decisions differ when some attributes of a tradeoff are presented numerically while others are presented qualitatively. Although numbers can be pallid and harder to process than stories or verbal descriptions (and low numeracy can reduce reliance on numbers), the authors propose that numbers confer a psychological advantage in tradeoff decisions because they are inherently suited for comparison (e.g., subtracting, adding, judging magnitudes). They hypothesize that decision-makers will overweight quantified attributes in comparative contexts due to greater comparison fluency—the felt and actual ease of using information to judge differences—leading to quantification fixation: a systematic preference shift toward options that dominate on the quantified dimension.

The work builds on several strands of research. Prior findings suggest numbers can be less vivid (e.g., anecdotal bias; sympathy for identifiable victims) and that numeracy affects processing and reliance on numeric information. Yet in tradeoffs, comparative judgments often focus on differences between attributes, which are naturally expressed numerically (Tversky; preference reversals; mindless math). Past research shows consumers value even superfluous metrics, sometimes ignoring what numbers omit. The authors connect their theorizing to evaluability theory (Hsee; General Evaluability Theory), which identifies mode (joint vs. separate), knowledge (distributional information), and nature (innate reference systems) as objective contributors to evaluability. They argue that subjective evaluability via comparison fluency is a missing fourth factor. Fluency research (Oppenheimer; Shah & Oppenheimer; Graf et al.) indicates that ease of processing influences cue weighting. They also consider and test alternative mechanisms (e.g., perceived importance signals, precision of numeric information, better encoding/recall, ambiguity aversion) and report limited or null roles for these compared to comparison fluency.

Across 21 preregistered experiments (8 main-text; 13 supplemental), participants made choices involving tradeoffs between two attributes; the authors randomized which attribute was quantified versus presented qualitatively (verbal descriptors, icon graphs, or continuous bar graphs). Objective evaluability was held constant via joint presentation, equivalent information, and explicit verbal–numeric mappings where relevant. Designs and samples included:

• Experiment 1a (N=1,000, Amazon Mechanical Turk): Hotel choice with price vs. rating tradeoff. Two conditions randomized whether rating or price was quantified; the other attribute was presented pictorially using icon bars (cash or stars). Ranges for price ($100–$500) and rating (1–5 stars) were given.
• SI S1: Separate evaluation variant of the hotel choice, presenting one hotel at a time and varying whether rating or price was quantified.
• SI S3b: Restaurant choice tradeoff (price vs. commute distance) using icon graphs; SI S3c: Assessed perceived importance of attributes to test a potential signaling mechanism; SI S3a: Property choice (annual property tax vs. distance to city center).
• Experiment 1b (N=1,000, Prolific): Summer internship candidate choice with calculus grade vs. management grade tradeoff. One grade shown as an imprecise numeric range (e.g., 93–97%), the other as a letter grade.
• SI S4: Tested the perceived precision of numeric descriptors by using ranges vs. point estimates; quantification fixation persisted.
• Experiment 1c (N=1,000, Prolific): Conference location choice (connectedness vs. sustainability). Both attributes had 1–5 scores with transparent verbal mappings to numeric values; randomized which attribute was quantified vs. verbal.
• Experiment 2 (N=2,000, Prolific): Employee promotion decision (likelihood of advancement vs. likelihood of retention). Four conditions: advancement quantified, retention quantified, both quantified, neither quantified. Verbal descriptors were explicitly mapped to numeric probabilities (e.g., “Almost certain = 95%”; “Likely = 70%”) and matched using prior probability word perception data. Analysis used OLS with robust SE, comparing selection rates across conditions.
• Experiment 3a (N=1,000, Prolific): Incentive-compatible hiring task. Candidates had scores on Math and Angles Games (0–10). Participants earned bonus based on the chosen candidate’s Trivia Game performance ($0.05 per correct). Scores were conveyed either numerically or via continuous bar graphs; randomized whether Math or Angles was quantified. Stimulus sampling included two candidate pairings with 3- or 4-point tradeoffs.
• Experiment 3b (N=701, in-person at three sites): Real $1 donation choice between The Natural Resources Defense Fund and The Nature Conservancy, each with Accountability & Finance and Culture & Community scores (from Charity Navigator). Randomized which attribute was quantified vs. bar graph; site fixed effects in regression.
• SI S7: Field replication via Meta ads directing users to vote for a $1,000 donation between the same charities.
• Experiment 4 (N=2,000, Amazon Mechanical Turk): Public works project choice (benefit to the community vs. efficiency). Randomized quantified attribute (benefit vs. efficiency) and numeric fluency. Fluent numbers (75/100; 25/100) vs. disfluent ratios (51/68; 23/92), validated in pretests for perceived comparison fluency. Analysis: OLS with robust SE, testing interaction of quantified attribute and fluency condition.
• SI S8: Replication with different numeric values (e.g., 90/100).
• SI S9a/S9b: Mediation tests using a three-item fluency measure (comfort, confidence, ease) comparing numeric vs. non-numeric information.
• Experiment 5 (N=602, Qualtrics Panels; nationally representative US adults): Real $1 donation choice scenario as in 3b. Measured objective numeracy (4-NUM; 4 items scored 0/1; M=1.25, SD=0.94) and subjective numeracy (8-item scale; Cronbach’s α=0.86; ability α=0.91; preference α=0.71). Explored moderation via OLS models with robust SE, interactions between condition assignment and numeracy measures. All experiments were preregistered, randomized via Qualtrics, and IRB-approved (UPenn ID 849979; UChicago Booth ID 23-1348). Data/materials/code: OSF (https://osf.io/97peh/?view_only=566b843cf41a46f9962237423078597e).

Core effect—quantification fixation: Across contexts, people favor options that dominate on the quantified attribute.

• Experiment 1a (hotel choice): Higher-rated/higher-price hotel (Hotel Luxe) chosen more when rating was quantified vs. when price was quantified: 51.6% vs. 33.0%; χ²(1)=34.68, P<0.001, 95% CI [0.124, 0.248], h=0.379. • SI S1 (separate evaluation): Significant interaction indicating quantification fixation; b_interaction=0.281, SE=0.062, 95% CI [0.160, 0.401], t(996)=4.56, P<0.001. • SI S3b (restaurant price vs. commute): Strong shift toward cheaper/longer commute when price quantified; χ²(1)=83.28, P<0.001, 95% CI [−0.315, −0.205], h=0.598. • SI S3c (attribute importance): No difference in perceived importance (cost vs. commute) by quantification condition; 72.4% vs. 76.4%, χ²(1)=0.851, P=0.356; equivalence test confirms outcomes equivalent at 90% confidence (±0.105 tolerance). • SI S3a (property tax vs. distance): χ²(1)=31.4, P<0.001, 95% CI [−0.208, −0.100], h=0.363.
• Experiment 1b (grades familiar numerically and non-numerically): Choosing candidate with higher management grade more when management grade quantified vs. calculus grade quantified: 83.8% vs. 68.9%; χ²(1)=29.94, P<0.001, 95% CI [0.095, 0.203], h=0.355. Quantification fixation persisted with imprecise numeric ranges; SI S4 showed persistence with range vs. point estimates.
• Experiment 1c (transparent verbal–numeric mapping): Higher connectedness/lower sustainability location chosen more when connectedness was quantified vs. sustainability quantified: 78.0% vs. 60.8%; χ²(1)=34.02, P<0.001, 95% CI [0.114, 0.230], h=0.377.
• Experiment 2 (format distortion): Baselines with consistent formats: both quantified=27.9% chose higher advancement/lower retention; neither quantified=32.7%; no significant difference (P=0.104). Quantifying one attribute distorted preferences: advancement quantified increased selection of higher advancement employee to 44.2% (P<0.001); retention quantified decreased it to 21.8% (P<0.05). Standard quantification fixation replicated: advancement quantified > retention quantified (P<0.001).
• Experiment 3a (incentive-compatible hiring): Higher Math/lower Angles candidate chosen more when Math quantified vs. Angles quantified: 66.5% vs. 54.5%; χ²(1)=14.46, P<0.001, 95% CI [−0.182, −0.057], h=0.245. Earnings: Mean bonus $0.28 (Math quantified) vs. $0.26 (Angles quantified); t(996.41)=3.49, P<0.001; 6% lower in Angles quantified despite Math being more predictive of Trivia performance.
• Experiment 3b (real in-person donations): Donated more to NRDC (higher Accountability & Finance, lower Culture & Community) when Accountability & Finance was quantified vs. Culture & Community quantified: 56.7% vs. 41.4%; b=0.153, SE=0.037, 95% CI [0.080, 0.226], t(697)=4.09, P<0.001. Field replication (SI S7): 48.3% vs. 35.6%; χ²(1)=3.41, P=0.065, 95% CI [−0.006, 0.260], h=0.258.
• Experiment 4 (mechanism—comparison fluency): Significant interaction indicating attenuation of quantification fixation when numbers were disfluent: b_BenefitQuantified×DisfluentNumber=−0.148, SE=0.037, 95% CI [−0.221, −0.075], t(1996)=−3.99, P<0.001. No main effect of benefit quantified (b=0.017, P=0.369) nor disfluent numbers (b=0.034, P=0.210). Replication SI S8: P=0.025. Mediation by fluency (SI S9a/S9b): comfort, confidence, ease partially mediate quantification fixation.
• Experiment 5 (moderation by numeracy; nationally representative): NRDC chosen more when Accountability & Finance quantified vs. Culture & Community quantified: 56.0% vs. 25.9%; χ²(1)=55.22, P<0.001, 95% CI [0.223, 0.379], h=0.623. Subjective numeracy moderated the effect: significant interaction in Model 2 (P=0.005) and Model 3 (P=0.010); objective numeracy showed no significant moderation (Model 1 P=0.196; Model 3 P=0.842). Result robust controlling demographics (P<0.001).

The findings establish that quantification is not neutral in tradeoff decisions: presenting one attribute numerically increases its comparative salience and weight, producing preference shifts toward options that dominate on the quantified dimension. This quantification fixation appears across managerial, policy, and consumer contexts, in joint and separate evaluations, and even with real incentives (earnings, donations) and in the field. The proposed mechanism—comparison fluency—explains why numbers shape decisions: numeric attributes are easier to compare, heightening sensitivity to magnitude differences and increasing reliance on quantified dimensions. Attenuating numeric comparison fluency (e.g., less interpretable ratios) weakens quantification fixation; individuals higher in subjective numeracy (greater comfort with numbers) show stronger fixation, and perceived fluency (comfort, confidence, ease) partially mediates the effect. These results extend evaluability theory by adding comparison fluency as a subjective component of evaluability that can skew decisions even when objective evaluability (mode, knowledge, nature) is held constant. Practically, as platforms and organizations increase quantification (e.g., ratings, prices), decision architectures may systematically pull attention to numeric attributes, potentially underweighting valuable qualitative information.

The paper demonstrates a robust decision-making distortion—quantification fixation—where decision-makers privilege quantified attributes in tradeoffs. It documents financial and behavioral consequences across diverse contexts and provides converging evidence that comparison fluency drives the effect. Contributions include: (1) identifying and naming quantification fixation; (2) showing its generality and incentive compatibility; (3) introducing comparison fluency as a key, subjective component of evaluability; and (4) clarifying boundary conditions via numeric disfluency and subjective numeracy. Future research should test interventions that increase the comparison fluency of qualitative information to reduce fixation, examine whether nonquantification neglect occurs (underweighting less fluent attributes), tease apart actual vs. felt ease contributions, and explore domains where numeric information is aversive or ethically fraught, potentially reversing the effect.

While experiments hold objective evaluability constant and show mediation by perceived fluency, the work does not isolate the relative roles of actual vs. felt ease; both likely contribute. Numbers were recalled marginally better than non-numbers, but this did not mediate the effect (weak mechanism). Some supplemental field recruitment was underpowered (SI S7), yielding marginal significance despite comparable effect sizes. The studies primarily involve short, simplified tradeoff scenarios; generalizability to complex multi-attribute, high-stakes domains warrants further testing. Potential alternative mechanisms (e.g., perceived trust in numeric information, ambiguity aversion to qualitative descriptors) were probed but may still play context-dependent roles. The mapping of author superscripts to affiliations for one author was not explicitly provided in the text.