Skip to content

Elasticity Experiment

Experimental method — instantiates Rebound-Aware Efficiency Governance

Deliberately tests several lever magnitudes, messages, or friction levels on small slices before scaling, to measure how strongly demand rebounds — the elasticity every price and guardrail is tuned against.

Elasticity Experiment is the archetype's instrument for measuring behavior rather than assuming it. Instead of guessing how much demand will bounce back when efficiency makes a service cheaper, it tests it: it runs several magnitudes of a lever — price levels, messages, friction settings — on small, bounded slices of the population, holds a slice back as a control, and reads the response before anyone scales the chosen setting. Its output is an elasticity — how strongly demand moves per unit of price or friction — together with a map of which behavioral channels the response travels down. This is the number the price levers and guardrails are all tuned against. Its defining move is deliberate, prospective variation: it manipulates the lever on purpose across arms, which is what lets it read cause and trace a forward-looking curve, rather than evaluate something already done.

Example

A utility has rolled out efficient heat pumps and is worried about take-back — that households, finding heat suddenly cheap, will simply keep their homes warmer and erode the saving. Before committing to a rebate level or a price signal across the whole service territory, it runs a winter-long randomized trial. Several thousand homes are split into arms: different rebate sizes, different "bank your savings" messages, and a holdout that gets nothing. Metered gas and electricity use are compared across arms. The result is not a single guess but a curve — how much take-back each rebate level and each message produces — plus a read on which channel dominates, whether households creep the thermostat setpoint up or run the heat for longer hours. That curve is what lets the eventual policy be set at a magnitude known to work, instead of one hoped to.

How it works

The method is design-of-experiment applied to demand. Multiple treatment arms span a range of lever magnitudes so the shape of the response — not just its sign — can be read; a control arm nets out weather, price, and trend; randomization (or the cleanest quasi-random assignment available) is what turns the measured difference into a causal estimate. Because rebound matures — take-back tends to grow as a cheaper habit settles in — the experiment's credibility rests on running long enough to catch the slow part of the response, not just the novelty spike.

Tuning parameters

  • Number and spacing of arms — how many lever magnitudes are tested and how far apart. More arms trace the curve's shape; fewer give only a point and a direction.
  • Sample size and power — how many units per arm. Under-powered arms cannot distinguish a real elasticity from noise, which is how weak experiments launder assumptions.
  • Duration — how long the trial runs. Short trials catch the novelty response; only a long one catches the rebound that grows as behavior settles.
  • Randomization vs quasi-experiment — true random assignment versus matching on observed traits. Randomization buys causal cleanliness; a quasi-design buys feasibility where randomizing is impossible.
  • Outcome measure — metered use versus self-report. Metering is honest but costly; self-report is cheap and systematically optimistic.

When it helps, and when it misleads

Its strength is that it replaces an assumed elasticity with a measured one, catches rebound before it is baked into a policy at full scale, and can separate the channels the take-back travels down so the fix can target the right one. Where the whole design of a price or guardrail hinges on how strongly demand answers, this is the mechanism that supplies the answer with evidence.

It misleads chiefly through external validity: a small, watched slice may not behave like the full rollout, and elasticity itself drifts with scale and time, so a clean trial number can quietly expire.[1] Novelty and observation effects inflate early response; short trials miss the long-run take-back that matters most; and it costs time a decision may not have. The classic misuse is to run a single arm, or to elevate the arm that confirms the desired policy, and call the result an elasticity. The discipline is multiple arms against a real holdout, a run long enough for rebound to mature, and a pre-committed outcome measure.

How it implements the components

  • elasticity_and_behavior_estimate — the primary output: a measured curve of how demand responds to the lever, the number the archetype's price and guardrail mechanisms consume.
  • response_path_map — the experiment reveals which behavioral paths the response travels, not just its size, by contrasting arms and messages.
  • rebound_channel_map — by instrumenting several outcomes it distinguishes the direct and indirect channels through which take-back arrives.

It does not recover the counterfactual for an intervention already deployed — that backward-looking evaluation is Control Group Comparison — and it sets no price itself; Price Incentive Adjustment and Demand Response Pricing consume its estimate.

References

[1] Price elasticity of demand — the percentage change in quantity demanded per percentage change in price. It is not a constant of nature: it varies across users, magnitudes, and time horizons, which is why an elasticity measured on a small short trial must be scaled and refreshed with care. ↩