Reference Dependent Utility over Prices and the Welfare Costs of Inflation

Last registered on August 04, 2026
View Trial History

Pre-Trial

Trial Information

General Information

Title
Reference Dependent Utility over Prices and the Welfare Costs of Inflation
RCT ID
AEARCTR-0018511
Initial registration date
July 31, 2026

Initial registration date is when the trial was registered.

It corresponds to when the registration was submitted to the Registry to be reviewed for publication.

First published
August 04, 2026, 9:58 AM EDT

First published corresponds to when the trial was first made public on the Registry after being reviewed.

Locations

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information

Primary Investigator

Name
SHARATH SONTI
Affiliation
Contact Primary Investigator

Other Primary Investigator(s)

Additional Trial Information

Status
In development
Start date
2026-08-15
End date
2026-09-30
Keywords
Behavior, Other, Welfare
Additional Keywords
Shopping, Price Expectations, Reference Prices, Reference-Dependent Preferences, Inflation Expectations, Welfare Costs of Inflation, Behavioral Macroeconomics.
JEL code(s)
C83, C91, D12, D84, D91, E31, D91, E71
Secondary IDs
Prior work
This trial does not extend or rely on any prior RCTs.
Abstract
Households dislike inflation more than standard economic models can explain. This project studies one reason rooted in consumer psychology: consumers carry an internal "reference price" for the goods they buy regularly, and a posted price above that reference feels like a loss. Built into an otherwise standard demand model, this mechanism makes inflation costly through a channel standard welfare accounting misses. Welfare losses from this mechanism are governed by two parameters: the elasticity of demand to changes in the reference price, holding the posted price fixed, and how slowly the reference price adjusts after prices change.

This online survey experiment, fielded on Prolific to a representative U.S. sample (ages 18-65), estimates both parameters. Each respondent first names a grocery or household product they buy regularly and their usual store, then reports the price they typically expect to pay for the product (our elicitation of the consumer's reference price for that product) and how many units they would buy over the next month at that price. Respondents then face two hypothetical scenarios built to move the reference price while holding the posted price fixed: the price the respondent actually pays next month is fixed at their originally elicited expected price, but the product's usual, long-run price is shifted away from it by a randomly assigned amount (12%, 20%, or 28%). In the "loss" scenario the usual price is set below the posted price, so the fixed posted price feels expensive relative to the scenario's usual price. In the "gain" scenario the usual price is set above the posted price, so the same posted price feels like a bargain. Because the posted price is identical across scenarios, any difference in how much the respondent says they would buy and in the price they report expecting to pay going forward (our elicitation of the shifted reference price) reflects a change in the reference price in the hypothetical scenarios. A separate block measures the speed of reference price updating. Respondents are told that the price of their product was at their initial reference price on previous shopping trips, but that on their next trip they encounter a price 20% higher. They are then asked, in one of two randomly assigned framings, what price they would expect to pay on subsequent trips (E arm) or what price would feel normal on subsequent trips (N arm). This elicitation captures the extent to which consumers' reference prices update in response to a higher posted price on one shopping trip. Because respondents also report how often they shop for the product, we can convert this trip-level updating speed into updating at the quarterly frequency of the model: a respondent's per-trip updating, compounded over the number of trips they take in a quarter, gives the quarterly reference price persistence that the welfare calculation requires.

We pre-register four confirmatory hypotheses: (1) demand responds to the reference-price gap in the loss scenario; (2) demand responds to the reference-price gap in the gain scenario; (3) the loss-side response is larger than the gain-side response; and (4) in both updating framings, reference prices adjust only partially after a single exposure to a higher price, rather than jumping immediately to the new price. This registration covers a new wave of approximately 1,000 respondents, to be pooled with approximately 1,300 respondents collected in April 2026 using the identical instrument; the pre-analysis plan fixes the estimation procedure, exclusion rules, and hypothesis tests before the new wave is fielded.
External Link(s)

Registration Citation

Citation
SONTI, SHARATH. 2026. "Reference Dependent Utility over Prices and the Welfare Costs of Inflation." AEA RCT Registry. August 04. https://doi.org/10.1257/rct.18511-1.0
Sponsors & Partners
Experimental Details

Interventions

Intervention(s)
The intervention consists of randomized manipulations embedded in an online survey; there is no field intervention and no treatment administered outside the survey instrument.

Each respondent first fixes the object of measurement: a grocery or household product they buy regularly, the package size in which they buy it, and their usual store. These are piped into every subsequent question, so all prices and quantities refer to a concrete, familiar good at a fixed unit. The respondent then reports the per-unit price they typically expect to pay for the product at that store -- the survey's elicitation of the consumer's reference price (R0) -- and the number of units they would buy over the next month at that price (the baseline quantity q0).

Randomized manipulation 1 -- reference-price shift size. Each respondent is randomly assigned a shift size s of 12%, 20%, or 28% with equal probability, held constant across both of their scenarios. The two scenarios are built to move the reference price while holding the posted price fixed: the price the respondent actually pays next month is fixed at their originally elicited reference price R0, while the product's usual, long-run price is shifted away from it by s. In the loss scenario the usual price is set below the posted price, at (1-s)R0, so the fixed posted price is expensive relative to the scenario's usual price; in the gain scenario the usual price is set above it, at (1+s)R0, so the same posted price is a bargain. Within each scenario the respondent reports (i) the number of units they would buy next month at the posted price and (ii) the price they would expect to pay going forward once the usual price resumes -- the elicitation of the shifted reference price.

Randomized manipulation 2 -- scenario order. Whether the loss or the gain scenario is shown first is randomized with equal probability.

Randomized manipulation 3 -- reference-price updating framing. In a separate block, the respondent is told that the price of their product was R0 on previous shopping trips and that on their next trip they encounter a price 20% higher. They are then asked what price they would carry forward to subsequent trips, in one of two randomly assigned framings: the expected-price arm (E) asks what price they would typically expect to pay going forward, while the feels-normal arm (N) asks what price would feel like the normal price going forward -- a price they have gotten used to paying that feels neither too high nor too low. The E arm captures an articulated price expectation; the N arm captures the internal evaluative benchmark against which posted prices are judged.

The survey closes with descriptive items: item perishability, shopping frequency for the product (used to convert per-trip updating to the model's quarterly frequency), reference-price feeling, salience, and formation channels, and emotional reactions to a surprise 20% price increase and a surprise 20% price decrease. These items are asked after the scenario quantity questions and are post-treatment relative to them.
Intervention Start Date
2026-08-15
Intervention End Date
2026-09-30

Primary Outcomes

Primary Outcomes (end points)
Survey outcomes: the loss-scenario purchase quantity (q_L); the gain-scenario purchase quantity (q_G); the elicited loss-scenario reference price (R_L); the elicited gain-scenario reference price (R_G); and the post-shock updated reference price (R_hat), elicited in either the expected-price (E) or the feels-normal (N) framing.

Derived estimands: the loss-side reference-dependence elasticity of demand (Gamma_L); the gain-side reference-dependence elasticity (Gamma_G); the loss-gain asymmetry (Delta_Gamma = Gamma_L - Gamma_G); and per-trip reference-price persistence (rho_trip), separately by updating arm.
Primary Outcomes (explanation)
The reference-dependence elasticities are constructed as follows. For respondent i in scenario j (loss or gain), the regressor is the log reference-price gap x_ij = log(R_ij / P_i), where P_i = R0 is the posted price held fixed across scenarios and R_ij is the elicited going-forward reference price. The within-respondent identification result implies log E[q_ij | x_ij] = log q0_i + Gamma * x_ij, so Gamma is estimated by Poisson pseudo-maximum-likelihood with no intercept and log q0_i as a fixed offset. Because the elicited reference R_ij is self-reported, x_ij is instrumented by the announced log gap implied by the randomly assigned shift size s, using a control-function approach: x is regressed on the announced gap by OLS and the residual enters the second-stage Poisson regression. Standard errors come from a respondent-clustered bootstrap with 1,000 replications.

Hypothesis 1 tests Gamma_L > 0 in a single-scenario loss specification; Hypothesis 2 tests Gamma_G > 0 in a single-scenario gain specification; Hypothesis 3 tests Delta_Gamma > 0 in a balanced specification stacking both scenarios (two observations per respondent, respondent-clustered). All three are one-sided Wald tests using bootstrap standard errors at alpha = 0.05.

Per-trip reference-price persistence is computed in closed form per respondent as rho_trip = (log R_hat - log p_shock) / (log P - log p_shock), where p_shock = 1.2 * R0 is the price shown in the updating block. rho_trip = 0 means the reference price jumps fully to the new price after one exposure; rho_trip = 1 means it does not move at all. Hypothesis 4 tests that mean rho_trip is strictly positive, by one-sample t-test run separately within the E and N arms at alpha = 0.05 each. No asymmetry test between arms is pre-registered. For the model calibration (not part of the confirmatory family), per-trip rho is compounded to a quarterly horizon using each respondent's reported shopping frequency and averaged within arm.

No multiple-testing adjustment is applied across the four hypotheses: each is a test on a conceptually distinct parameter. The confirmatory family is exactly these four tests, and no hypotheses will be added to it after this registration is filed. The exact estimating equations, the pre-specified exclusion restrictions, and the reference-price window defining the estimation samples are in the attached pre-analysis plan (Sections 5-7).

Secondary Outcomes

Secondary Outcomes (end points)
Reference-price feeling response (whether a posted price equal to the elicited reference feels too low, about right, or too high); post-shock feeling response (whether the higher price feels about right after persisting for a month); reference-price salience; reported reference-price formation channels; item perishability; shopping frequency; and emotional reactions to a surprise 20% price increase and to a surprise 20% price decrease.
Secondary Outcomes (explanation)
These items characterize how the reference price is held, formed, and used, and how respondents react to surprise price changes. None enters the confirmatory hypothesis family. Alongside them, a standard set of sensitivity tables already implemented in the analysis workflow will be reported: a pooled regression imposing a single reference-dependence elasticity with no loss/gain asymmetry; an OLS/2SLS counterpart to the full set of tables; and a demographics-by-wave balance table extended to include the new wave. Analyses using the diagnostic, item-characteristic, or emotional-response variables are exploratory.

Experimental Design

Experimental Design
Online survey experiment fielded on Prolific to a representative U.S. sample, ages 18-65. Randomization is at the individual respondent level, carried out independently by the survey software, and is neither clustered nor stratified. Three independent randomizations are applied to each respondent: a reference-price shift size of 12%, 20%, or 28% with equal probability; the order in which the loss and gain scenarios are shown, with equal probability; and assignment to the expected-price (E) or feels-normal (N) updating framing, with equal probability. Every respondent sees both the loss and the gain scenario, so the loss/gain contrast is within-respondent.

Target sample size is 1,000 collected responses. Data collection stops at exactly 1,000 collected responses regardless of interim results. No interim analyses will be performed and no data-dependent stopping is permitted. If Prolific over-recruits by a few responses, as commonly occurs, the final N at the moment collection is paused is taken as the analysis sample; we do not subsample down. Expected retention after the pre-specified baseline exclusions is approximately 85%, giving roughly 850 respondents eligible for analysis.

Disclosure of existing data. This registration covers an additional wave that extends an existing dataset of 1,323 respondents collected April 24-27, 2026 (1,262 after baseline exclusions) using the identical instrument, on which the reference-dependence elasticities and updating speeds have already been estimated. The primary specification pools the new wave with that prior data into a single combined sample with no wave indicator, giving an expected combined analysis sample of approximately 2,100 respondents and a balanced estimation sample of approximately 1,700. Pooling is justified ex ante on three grounds set out in the pre-analysis plan: identification is within-respondent, so cross-wave composition differences do not bias the estimated elasticity unless they correlate with the structural parameter itself; the prior waves were demographically balanced on age, sex, ethnicity, employment, student status, and lifetime Prolific approvals; and any level difference between waves shifts a respondent's baseline quantity, which the offset absorbs respondent by respondent. As an appendix robustness check -- not part of the primary strategy -- the pooled specification is augmented with a wave indicator and its interaction with the reference-price gap. A significant wave term will be reported and discussed but will not, by itself, trigger a deviation from the pooled-primary framing.

The prior-wave point estimates are reported in the attached pre-analysis plan (Section 8), placed deliberately after the specifications and exclusion restrictions so that the pre-registered analysis choices are stated before those estimates are reviewed. They are used for design and precision planning only. The purpose of this registration is to eliminate analyst degrees of freedom in the confirmatory tests going forward.
Experimental Design Details
Not available
Randomization Method
Randomization is done by the survey software (Qualtrics) at the moment a respondent enters the survey, using its built-in randomizer. The three randomized components (reference-price shift size, scenario order, and updating framing) are assigned independently of one another, each with equal probability. No stratification, no blocking, no clustering.
Randomization Unit
Individual survey respondent. All three randomized components are assigned at the individual level; there is no higher level of randomization.
Was the treatment clustered?
No

Experiment Characteristics

Sample size: planned number of clusters
Not clustered; randomization is at the individual level. Equivalently, 1,000 clusters of one respondent each (the new wave's 1,000 individual respondents).
Sample size: planned number of observations
1,000 collected survey responses in the new wave, with approximately 850 expected to remain after the pre-specified baseline exclusions. Pooled with the prior waves collected in April 2026 (1,323 collected; 1,262 after exclusions), the combined analysis sample is expected to be approximately 2,100 respondents, of whom approximately 1,700 enter the balanced specification with two scenario observations (one loss, one gain) each.
Sample size (or number of clusters) by treatment arms
New wave, before exclusions: approximately 333 respondents per reference-price shift size (12%, 20%, 28%); approximately 500 respondents per scenario order (loss scenario first vs. gain scenario first); approximately 500 respondents per updating framing (expected-price E vs. feels-normal N); approximately 83 respondents per full factorial cell (3 shift sizes x 2 orders x 2 framings = 12 cells). Every respondent sees both the loss and the gain scenario, so the loss/gain contrast is within-respondent rather than between arms.

Minimum detectable effect size for main outcomes (accounting for sample design and clustering)
Supporting Documents and Materials
IRB

Institutional Review Boards (IRBs)

IRB Name
Committee for Protection of Human Subjects (CPHS), University of California, Berkeley
IRB Approval Date
2026-04-21
IRB Approval Number
2026-02-19386
Analysis Plan

Analysis Plan Documents

Reference-Dependent Demand and Reference Price Persistence: Pre-Analysis Plan for an Additional Survey Wave

MD5: ce4643a21e40a8035d11249e8ad87923

SHA1: 1b71d8149162cb2d6b12d88dee95d25ceb8417d9

Uploaded At: July 31, 2026