The Lewsearch Report/Pulse
An extra $100 saving shifted simulated preference toward connecting flights
· Exploratory synthetic-panel experiment · Not graded
Would four extra travel hours feel worthwhile for $50 in savings—or $150?
In a paired hypothetical airfare test, reducing a connecting itinerary’s round-trip price from $450 to $350, while holding a nonstop alternative at $500, increased the weighted model-elicited probability of choosing the connection from 45.39% to 50.46%: a 5.07-percentage-point increase among profiles with valid responses to both offers.
These are simulated choice probabilities, not percentages of actual travelers.
The question is timely: BLS’s October 14 release reported that airline fares were 22.4% higher than a year earlier in September 2026, and rose 1.2% that month after seasonal adjustment. Those figures provide editorial context; they were not shown to the model in this experiment. BLS September 2026 CPI release.
The trade-off
Each fictional adult considered one self-paid ticket for a domestic holiday visit:
- Nonstop: $500 round trip; three hours each way.
- Connecting: $450 or $350 round trip; five hours each way, including the connection.
Destination, dates, comparable departure times, baggage and ticket-change terms stayed the same. Prices included taxes and fees. “Neither offer” and “not sure” remained available.
All four responses
| Option | Connection costs $450 | Connection costs $350 |
|---|---|---|
| $500 nonstop | 50.75% | 41.17% |
| Connecting itinerary | 45.39% | 50.46% |
| Neither offer | 3.86% | 8.37% |
| Not sure | 0.00% | <0.01% |
The cheaper connection gained probability, but “neither” also increased. This is not a simple one-for-one shift from nonstop to connecting flights, and “neither” should not be read as canceled travel.
How much does missing output matter?
We scheduled 500 synthetic profiles and 1,000 paired responses. Thirty-five responses failed the required probability-total check and were retained as missing, not repaired. The 467 complete pairs represent 96.04% of the original panel weight.
Allowing every missing response any probability from zero to one gives a full-scheduled-panel connecting-flight difference between +3.02 and +7.36 percentage points. Thus the direction survives this missing-output check. These are worst-case missingness bounds, not confidence intervals; they do not cover model error or repeat-run variability.
Method and limits
The 500 profiles were selected once from an existing 3,000-profile nationally structured synthetic roster, stratified across 20 region–age cells. Historical demographic weights were adjusted for subset selection. This is not a representative U.S. consumer poll: demographic mismatches remain, including approximately 4.16 percentage points for gender and 5.35 for party relative to the configured targets.
Each profile received the same context in both price conditions, with balanced offer order and fresh model contexts. Context included three stored synthetic belief states dated August 17–October 1—not refreshed news beliefs or invented travel preferences. Income descriptions used coarsened legacy ranges; the upper ranges overlap because original boundaries were lost. Historical income weights are approximate coded-category calibrations, not exact population shares for the overlapping rendered ranges. These beliefs are not necessarily airfare-relevant, article-linked or current.
Lewis-v17 produced verbalized four-option probabilities, once per profile and condition. An initial eight-call format attempt was kept separate. After a schema correction, the strict format gate failed again; a documented exploratory continuation authorized finishing the unchanged 1,000-request measurement set. There were 965 valid responses: 487 in the $450 condition and 478 in the $350 condition. No failed vector was rescaled or replaced.
Full response accounting
All 1,000 scheduled schema-corrected requests were attempted once: 500 per condition. Invalid outputs numbered 13 at $450 and 22 at $350. Of the 500 profiles, 467 had two valid responses, 11 had only the $350 response valid, 20 had only the $450 response valid, and two had neither valid.
The final measurement used temperature 1 and at most 64 completion tokens. No probability floor, temperature transform, response repair or fallback estimator was applied. Positive missingness bounds do not establish robustness to stochastic generation, alternative prompts, model error or human validation.
The takeaway: increasing the discount shifted this model-based panel toward the longer itinerary. It is a useful hypothetical price-sensitivity result—not a forecast of bookings, proof of human preferences or a replicated estimate.
Request research materials. No individual dossiers, attributes, prompts or row-level outputs are published here.