Experimental economics tests economic ideas by deliberately varying rules, information or incentives and observing the resulting choices. Its gold standard is an experiment whose question, treatment, comparison, payment rules, assignment procedure and limits are clear enough for another researcher to inspect and reproduce. A successful experiment can reveal how an institution works, challenge a theory or improve a practical decision; it does not make every result universal.
The central skill is to connect a controlled change with a defensible conclusion. Paying participants does not automatically reveal their true preferences, random assignment does not eliminate every source of error, and a laboratory result does not automatically describe a whole economy. This guide uses original fictional markets and numerical examples to make those distinctions visible. None of the teaching figures reports an actual study or programme.
Start with the Gold Standard master, the Education parent and the How X Works hub. The Behavioural Economics root examines patterns of choice; the Econometrics root develops measurement and inference. Here the focus is how to build and interpret an economic experiment.
Choose your experimental economics route
Build a clear comparison, calculate the incentives and judge what the resulting evidence can establish.
Route 1 · Chapters 1–3: Frame the question and incentives
Route 2 · Chapters 4–6: Calculate markets and social dilemmas
Route 3 · Chapters 7–9: Design and estimate comparisons
Route 4 · Chapters 10–12: Protect evidence and participants
Route 5 · Chapters 13–14: Transfer and continue learning
Chapter index
Frame the question and incentives · Chapters 1–3
1. Turn an economic question into a controlled comparison
Calculate markets and social dilemmas · Chapters 4–6
4. Calculate a market's gains from trade
Design and estimate comparisons · Chapters 7–9
7. Randomise the unit that receives the treatment
Protect evidence and participants · Chapters 10–12
10. Diagnose missing outcomes, spillovers and implementation failures
Transfer and continue learning · Chapters 13–14
CHAPTER 1 OF 14 · Frame the question and incentives
1. Turn an economic question into a controlled comparison
An economic experiment begins with something people can choose and a condition the researcher can change. Participants might buy, sell, contribute to a shared account, choose between risky prospects or decide whether to complete a task. The researcher specifies the available actions and observes what happens under different conditions.
Suppose a market has buyers, sellers and a rule for accepting offers. Asking whether the market is efficient is incomplete. Efficient relative to which feasible allocation, with what information and under which trading rules? A useful question might ask whether displaying standing offers increases the proportion of available gains from trade realised during a fixed trading period.
That question names a treatment, an outcome and a comparison. It also suggests what must remain stable: the distribution of buyer values, seller costs, trading time and participant instructions. If the visible-offer market also receives twice as much time, a difference cannot be attributed to offer visibility alone.
Separate a prediction from a purpose
A prediction says what a model expects to happen. A purpose says why the result matters. A model might predict truthful bidding under specified auction rules. A practical purpose might be to choose an allocation procedure that participants can understand. Truthful bidding and understandable participation are related but separate outcomes.
Write the prediction before observing results. For example: visible offers will increase realised surplus, measured as a fraction of the maximum feasible surplus, across independently assigned markets. State what evidence would count against it. A result that merely feels consistent with a broad story is a weak test.
The Nobel presentation of Vernon Smith's contribution explains the importance of laboratory markets for studying alternative trading institutions. The important teaching move is to make the institution itself a variable that can be compared, rather than treating market rules as an invisible background.
Before proceeding, try to describe your experiment without using the word better. Replace it with an outcome, unit, period and comparator. If you cannot do that, the first repair belongs in the question rather than in the statistical software.
CHAPTER 2 OF 14 · Frame the question and incentives
2. Understand what the laboratory controls
A laboratory can specify the action set, information, matching procedure, timing and payment formula. It cannot directly set a person's beliefs, attention, social concerns or understanding. Experimental control therefore means control over defined features of the environment, supported by checks on how participants encounter them.
Imagine two participants offered the same menu of choices. One understands the instructions; the other thinks a hypothetical payoff will actually be paid. Their screens are identical but their perceived decisions are different. Saving the screen design alone would not document the experiment adequately.
Choose the environment for the question
A simple laboratory market is useful when the aim is to isolate a trading rule. A field experiment may be useful when the aim concerns decisions made within an existing organisation or community. Neither setting wins automatically. Familiar context can improve relevance while introducing variation that needs to be measured.
Harrison and List's Field Experiments distinguishes settings through features including participants, information, commodities, tasks, stakes and environment. That framework helps a reader ask what changed between settings instead of reducing the comparison to artificial versus real.
For our market question, a controlled environment could assign numerical values to a fictional good. A later study might recruit people with trading experience or use a familiar commodity. Changing several features simultaneously can test whether a finding travels as a package, but it will not identify which changed feature explains a different result.
Prepare a short environment record: who could participate, what they knew, what they could do, what they were paid, whom they interacted with and what they believed would happen next. Record the exact instructions and any verbal clarifications. An improvised explanation given to only one session can become an unintended treatment.
The learner's check is simple: identify one condition genuinely controlled by the researcher and one condition only measured or assumed. This distinction prevents the phrase controlled experiment from becoming a claim that every relevant influence has been eliminated.
CHAPTER 3 OF 14 · Frame the question and incentives
3. Make incentives understandable and consequential
An incentive rule maps actions and outcomes into consequences. In an induced-value trading exercise, a buyer might receive an assigned redemption value and earn that value minus the price paid. A seller might receive an assigned cost and earn the selling price minus that cost. These rules create known monetary opportunities inside the experiment.
Suppose a buyer's value is 20 tokens and the purchase price is 12. The buyer earns 8 tokens. A seller with a cost of 7 earns 5 tokens at the same price. Their combined gain is 13 tokens: 20 minus 7. Price divides this gain between them; it does not change the combined gain for that particular trade.
A payment rule is part of the treatment
Now suppose each token converts to a stated amount of money. Participants need to understand the conversion, whether every round is paid, whether any round is selected for payment, and whether a participation payment is separate. A rule that is clear to the researcher may still be confusing to the person making the decision.
Use a comprehension question before consequential decisions: if your value is 20 and you pay 12, what do you earn? Ask about a second example with different numbers so that a memorised answer is insufficient. Define in advance how misunderstandings will be handled; selectively excluding inconvenient choices after seeing outcomes changes the sample.
Monetary rewards can make a decision consequential without erasing other motives. Someone may care about fairness, avoid embarrassment, enjoy competition or misunderstand the task. Therefore, a known monetary payoff is not the same as a complete description of subjective utility.
For an original diagnostic, imagine raising the token conversion while leaving the instructions confusing. A change in behaviour could reflect stronger attention, increased concern about losses or something else. A larger reward does not by itself establish which explanation is correct. Record understanding and consider a design that varies clarity separately from stakes.
The practical standard is traceability: every paid amount should be reconstructible from the announced rule and recorded choices. Reconcile totals before payment, keep personal payment information separate from research identifiers, and document any software or administrative exception.
CHAPTER 4 OF 14 · Calculate markets and social dilemmas
4. Calculate a market's gains from trade
Consider a fictional market with four buyers, each able to buy one unit, and four sellers, each able to sell one unit. Buyer values are 20, 16, 12 and 8 tokens. Seller costs are 3, 7, 11 and 15 tokens. Assume identical goods, no transaction costs and no effects on anyone outside the market.
Sort buyers from highest value to lowest and sellers from lowest cost to highest. The possible marginal gains are 20 − 3 = 17, 16 − 7 = 9, 12 − 11 = 1 and 8 − 15 = −7. The first three trades add value; the fourth would destroy value under these assumptions.
The efficient benchmark is 27 tokens
Maximum feasible surplus is 17 + 9 + 1 = 27 tokens. This benchmark is calculated from the assigned values and costs, not from a participant survey. It tells us what the experiment makes possible under the stated constraints.
At a common price of 11.5 tokens, the three highest-value buyers are willing to buy and the three lowest-cost sellers are willing to sell. The remaining buyer values the good at only 8, and the remaining seller requires at least 15. Buyer surplus totals 48 − 34.5 = 13.5. Seller surplus totals 34.5 − 21 = 13.5. Together they recover 27.
Now imagine only the two highest-gain trades occur. Realised surplus is 26 and allocative efficiency is 26/27, about 96.3%. Trade volume is only two-thirds of the efficient three-trade quantity. A volume measure and a surplus measure therefore tell different stories: the missed third trade carries only one token of surplus.
If instead the lowest-value buyer replaces the buyer with value 12 while the same three sellers trade, realised surplus falls to 20 + 16 + 8 − 3 − 7 − 11 = 23. Efficiency is about 85.2%. Counting three transactions alone would miss the allocation error.
Do not present these percentages as a complete welfare judgement. They omit distributional concerns, participation burdens and external effects by construction. The experiment's benchmark is useful precisely because its boundaries are explicit. When a real decision contains additional costs or objectives, add them to the evaluation rather than silently assuming the token surplus captures everything important.
CHAPTER 5 OF 14 · Calculate markets and social dilemmas
5. Compare institutions without assuming their outcomes
A trading institution specifies who may submit offers, who sees them, how agreement occurs and how payments are determined. Experimental economics can compare institutions while holding the underlying economic opportunities stable. The mechanism-design specialist article explains why rules and private information must be considered together.
For a simple auction exercise, assume two bidders have private values of 18 and 14 tokens for one item. Under a second-price sealed-bid rule, the highest bidder wins and pays the other bid. With truthful bids of 18 and 14, the winner pays 14 and receives a monetary surplus of 4.
Check the strategic comparison
Fix the other bid at 14. The bidder with value 18 can bid above 14 and win at a price of 14. Raising the bid further does not improve that payment. Bidding below 14 loses a profitable opportunity. More generally, under the standard private-value setting, no collusion and the specified payment rule, truthful bidding is a weakly dominant strategy. Tie handling should be stated explicitly.
Change the institution to a first-price auction. If bids are 12 and 11, the first bidder wins, pays 12 and receives 6. These are illustrative bids, not a calculated equilibrium. Predicting first-price bids requires assumptions about beliefs, values and preferences that the example has not supplied.
It would be wrong to use these two numerical outcomes to conclude that first-price auctions always raise less revenue. The observed bid pattern was chosen for illustration. A defensible experiment needs a prespecified comparison across independently assigned units, enough information about the value environment and suitable analysis of uncertainty.
Also separate mistakes from preferences. An overbid may reflect misunderstanding, a different interpretation of winning or another motive. A comprehension check can identify some problems, but it cannot read a participant's mind. Preserve the behaviour and state which interpretation the evidence supports.
A useful extension is to compare a rule after participants have practised it with the same rule on first exposure. That asks whether usability improves with experience. It also introduces learning and order effects, so the sequence itself must be designed rather than left to convenience.
CHAPTER 6 OF 14 · Calculate markets and social dilemmas
6. See a social dilemma in the payoff arithmetic
Experiments can examine how people respond when private monetary incentives differ from a group's monetary interest. Consider four participants, each endowed with 10 tokens. Each privately chooses a contribution between zero and ten to a common account. The account is multiplied by 1.6 and divided equally among all four participants.
Participant i's payoff is 10 minus their own contribution, plus 0.4 times the total contribution. The factor 0.4 is 1.6 divided by four. Every extra token contributed returns 0.4 to its contributor and 0.4 to each of the other three people.
Compare the individual and group changes
Holding others' contributions fixed, increasing your own contribution by one token changes your payoff by −1 + 0.4 = −0.6. Across the whole group, the same contribution changes total payoff by −1 + 1.6 = +0.6. The private and collective monetary incentives point in different directions.
If nobody contributes, each keeps 10 and the group total is 40. If everybody contributes 10, the common account contains 40, becomes 64 and pays 16 to each person. Group earnings rise to 64.
If three people contribute 10 and the fourth contributes nothing, the account becomes 48 and each receives 12 from it. Each contributor earns 12; the non-contributor keeps 10 as well and earns 22. The group total is 58. This calculation explains a temptation to free-ride without asserting that every person will take it.
The social-cooperation specialist article connects these dilemmas to reciprocity, reputation and institutions. Our exercise isolates a monetary structure. Actual contributions may reflect social preferences, beliefs about others, learning or confusion, and a single contribution cannot distinguish all these possibilities.
Changing anonymity, communication or repeated matching changes the experiment. A participant who expects future interaction may face incentives absent from a one-shot version. Keep the description of the game aligned with the inference. A classroom demonstration can reveal the arithmetic and stimulate discussion; it is not automatically evidence about a community's moral character or a national culture.
CHAPTER 7 OF 14 · Design and estimate comparisons
7. Randomise the unit that receives the treatment
Random assignment uses a known chance mechanism to allocate treatment. It makes the assignment independent of potential outcomes by design, rather than relying on people to choose their group. It does not guarantee that every realised sample has identical characteristics across groups.
The unit matters. If a trading rule applies to an entire market, the independently assigned unit is the market. Twenty traders sharing one set of rules do not create twenty independent assignments of that rule. Repeated rounds within the same market also share participants and experience.
Match the analysis to the assignment
Suppose twelve separate markets are randomly divided between visible and hidden offers. Record the market identifier, assignment, session, participants and round. Analyse dependence at the level justified by the design. Treating every transaction as an independent trial can make the evidence appear much more precise than it is.
The J-PAL randomization guide discusses assignment methods, stratification and spillovers. Stratification means assigning within predefined groups, such as session times or relevant baseline categories, to balance an important feature through the design. It is different from repeatedly drawing assignments until an appealing result appears.
For an original planning exercise, list eight markets in four pairs with similar scheduled participant experience. Randomly assign one market in each pair to each institution. Preserve the pairing in the analysis. This design does not remove all uncertainty, but it makes the intended comparison explicit and limits a particular source of imbalance.
Random assignment is also different from random sampling. Volunteers can be assigned randomly within a study even when they are not a representative sample of the population. The treatment comparison may be credible for those participants while broader generalisation remains uncertain.
Record the eligible units before assignment, the actual randomisation procedure and the realised allocation. Protect allocation information where early knowledge could influence recruitment or participation. If someone overrides an assignment, preserve the original allocation and record the deviation. Rewriting the allocation history to match what happened would conceal the very problem the analysis needs to examine.
CHAPTER 8 OF 14 · Design and estimate comparisons
8. Estimate an effect and show its uncertainty
Consider a separate fictional experiment with 200 independently assigned people. Half receive a reminder and half receive no reminder. The prespecified outcome is completion of a task within seven days, and outcomes are observed for all participants. Assume no interference between people and no other difference introduced by the assignment.
Sixty of the 100 people assigned a reminder complete the task; 45 of the 100 assigned no reminder complete it. The estimated intention-to-treat effect is 0.60 − 0.45 = 0.15, or 15 percentage points. Relative to the control rate, that is a 33.3% increase. These two descriptions use different denominators and should not be confused.
Work through an approximate interval
For a simple independent two-proportion approximation, the standard error is the square root of [0.60 × 0.40/100 + 0.45 × 0.55/100]. The quantity inside the square root is 0.004875, giving a standard error of about 0.0698, or 6.98 percentage points.
An approximate 95% normal interval is 0.15 plus or minus 1.96 × 0.0698. Its endpoints are about 0.013 and 0.287: an increase of roughly 1.3 to 28.7 percentage points. This interval is deliberately wide. The point estimate is not a promise that another group will experience exactly 15 additional completions per hundred people.
This is a teaching approximation, not a universal analysis recipe. Clustered assignment, small samples, unequal probabilities, repeated observations or a different prespecified design require appropriate methods. A randomisation-based analysis should reproduce the actual assignment mechanism when constructing its reference distribution.
An intention-to-treat comparison keeps people in their assigned groups whether or not they read the reminder. Comparing only people who opened it with people who did not introduces a choice that may be related to motivation. The effect of offering a reminder and the effect of actually reading it are distinct questions.
The economists' causal-evidence guide provides a wider route into comparison and interpretation. Report the outcome definition, group counts, rates, effect scale and uncertainty together. A bare claim that the intervention worked hides information a decision-maker needs.
CHAPTER 9 OF 14 · Design and estimate comparisons
9. Plan precision, outcomes and interactions in advance
An experiment should be designed to learn something useful at the scale that matters. A sample can be too small to distinguish a practically relevant effect from noise. A statistically detectable effect can also be too small to justify the intervention's cost or burden.
The J-PAL power-calculation resource explains how power depends on sample size, variability, effect size, allocation and clustering. Planning requires assumptions, so examine a range of plausible inputs rather than treating one calculation as a guarantee.
Repeated observations do not always add independent information
As a rough illustration for equal-sized clusters, the familiar design-effect expression is 1 + (m − 1)ρ, where m is cluster size and ρ is the within-cluster correlation. With twenty observations per cluster and ρ = 0.05, the expression is 1.95. Two hundred clustered observations would then have a rough variance-equivalent count near 200/1.95, or 103 independent observations under that simplified model.
This does not make ten clusters equivalent to 103 independent randomised units. The number of clusters, degrees of freedom and exact design still matter. The calculation is a warning about dependence, not permission to ignore it.
Now consider a fictional two-by-two design testing a reminder and a simpler form. Completion rates are 40% with neither, 50% with the reminder only, 44% with simplification only and 62% with both. The reminder effect without simplification is 10 percentage points; with simplification it is 18. The additive interaction is 62 − 50 − 44 + 40 = 8 percentage points.
Those cell means demonstrate how an interaction is calculated. They do not establish that it is reliably different from zero; that requires sample sizes and uncertainty. Nor does a significant effect in one subgroup and an insignificant effect in another establish a significant difference between the two effects.
Define primary outcomes, comparisons and interaction tests before examining results. Additional exploration can be valuable when labelled as exploration. Searching many outcomes and reporting only the most attractive result makes an apparently decisive finding much less informative than the reader is led to believe.
CHAPTER 10 OF 14 · Protect evidence and participants
10. Diagnose missing outcomes, spillovers and implementation failures
Random assignment can be sound while follow-up measurement fails. Suppose 100 people are assigned treatment, but outcomes are observed for only 60; forty of those observed complete the task. All 100 controls are observed, with 45 completions. The observed treatment rate is 40/60, about 66.7%, compared with 45% among controls.
It is tempting to report a gain of 21.7 percentage points. But the unobserved treatment participants may differ systematically from those followed up. The comparison has lost the simple protection supplied by observing everyone in their assigned group.
Bound what the missing outcomes could change
For this binary outcome, the treatment group's full-sample completion rate could be as low as 40%, if none of the forty missing people completed, or as high as 80%, if all did. With the control rate fixed at 45%, the effect lies between −5 and +35 percentage points under these extreme assignments of missing outcomes.
These are worst-case missing-data bounds for the illustrated counts, not confidence limits. They expose the information gap without assuming it away. Additional justified assumptions or follow-up evidence may narrow the range, but those additions should be visible.
J-PAL's data-analysis guidance discusses attrition and threats arising when comparison units are indirectly affected by treatment. If reminder recipients help untreated friends complete the task, the control condition includes exposure through friendship. The observed contrast then needs a different interpretation from isolated treatment versus no exposure.
Implementation logs help distinguish failure to deliver from lack of response. Record messages sent, delivery failures, changes in the platform and the timing of outages. Do not redefine the primary group after discovering which recipients were easiest to reach.
A practical repair sequence is to recover missing outcomes through a procedure applied consistently, explain what remains missing, examine sensitivity and narrow the conclusion if necessary. A new regression term cannot automatically restore information that was never collected. When evidence cannot support the original question, reporting the limitation is part of completing the analysis.
CHAPTER 11 OF 14 · Protect evidence and participants
11. Protect participants and the credibility of the task
Economic experiments involve people interpreting instructions, making choices and trusting payment promises. Respectful treatment is part of the research design. Participation should be genuinely voluntary within the applicable review framework, with appropriate information, privacy protection and a clear account of payments and withdrawal arrangements.
Avoid treating a classroom or workplace as a convenient pool whose members have no practical choice. Authority relationships can affect both participation and behaviour. For learning activities, distinguish an optional demonstration from research intended to produce generalisable evidence, and obtain appropriate institutional guidance before collecting research data.
Do not confuse secrecy with a sound design
Experimental economics has a strong norm against deception. Charness, Samek and van de Ven's study also shows that researchers and participants can disagree about ambiguous practices. A procedure can create a misleading impression even when no single sentence is literally false.
For our trading exercise, do not describe a computer-controlled counterpart as another participant. If some information must be withheld for a justified design, the ethical review and participant information should address that procedure. This guide is not a substitute for the relevant review process.
Demand effects are another concern: participants may infer what the researcher hopes to observe and respond to that perceived expectation. De Quidt, Haushofer and Roth study methods for assessing sensitivity to experimenter demand. Their work supports investigating the problem rather than assuming that paying participants removes it.
An original warning example is an instruction saying that generous contributors help the study succeed. In the public-goods task, this wording introduces an appeal that was absent from the monetary model. If only one treatment receives that appeal, the comparison contains both the intended institutional difference and the wording difference.
Use neutral, tested instructions and preserve their exact versions. Keep comprehension assistance consistent. Record who could see individual choices and whether the experimenter interacted differently across groups. These steps improve interpretability, but they cannot certify that all expectations or social pressures have vanished. Describe what was done and what uncertainty remains.
CHAPTER 12 OF 14 · Protect evidence and participants
12. Register, reproduce and learn from replication
A durable experiment record should make the sequence of decisions visible. Preserve the original question, design, assignment procedure, instructions, outcome definitions and analysis plan before results are inspected. Later changes may be necessary, but their timing and reasons matter.
The AEA RCT Registry guidelines describe registration fields including outcomes, design, clustering and planned observations. Registry review does not certify scientific quality or ethical acceptability. Registration is evidence of what was recorded and when, not a stamp that makes the design correct.
Build a reproducible path from choices to results
Keep raw observations separate from cleaned data. A cleaning script should explain exclusions, duplicate records, payment conversions and construction of outcomes. Preserve identifiers that connect transactions to rounds, participants and markets without exposing unnecessary personal information.
For the fictional market, a second person should be able to recover the 27-token feasible-surplus benchmark from the assigned values, calculate realised surplus from accepted trades and reproduce the efficiency ratio. If the final table requires undocumented manual changes, the analysis is not yet reproducible.
Replication asks a further question: does a new implementation produce evidence consistent with the original claim? Camerer and colleagues' replication study examined eighteen laboratory economics studies and found a statistically significant effect in the original direction in eleven. That finding concerns the selected studies and the stated criterion; it is not a universal replication rate for economics.
A replication can differ because of sampling uncertainty, implementation, population, context or an unstable original estimate. Examine effect sizes and intervals as well as a binary significance label. Two studies can have similar estimates yet land on different sides of a threshold because their precision differs.
Publish null and inconvenient results with the same care as favourable ones. Keep exploratory findings distinguishable from prespecified tests. A useful report allows a reader to see which parts are observations, which are calculations, which are assumptions and which are proposed explanations still needing a new test.
CHAPTER 13 OF 14 · Transfer and continue learning
13. Test transfer when the group changes
Return to the four-person public-goods exercise. Its contribution multiplier was 1.6, giving each member 0.4 of every contributed token. Suppose the next exercise has eight participants while retaining the same multiplier and the same ten-token endowment per person. What changes?
The individual return from a contributed token becomes 1.6/8 = 0.2. Holding everyone else's contribution fixed, contributing an extra token changes the contributor's monetary payoff by −0.8 rather than −0.6. Across the group, the net gain remains +0.6 per additional contributed token.
Check the answer before carrying over a prediction
If all eight contribute ten, the account contains 80, becomes 128 and pays 16 to each. If nobody contributes, each earns ten. The full-cooperation payoff per person is unchanged even though the individual temptation to withhold a token is stronger under the monetary calculation.
If seven contribute ten and one contributes nothing, the common return per person is 112/8 = 14. Each contributor earns fourteen; the non-contributor earns twenty-four. The arithmetic is different from the four-person case even though both tasks are called a public-goods game.
Would contributions necessarily fall? The monetary model identifies one changed incentive, but the behavioural answer is not guaranteed. Group size may also alter perceived responsibility, beliefs, attention or social expectations. A comparison changing group size and participant population at the same time cannot isolate group size alone.
This is the transfer habit emphasised in the Education parent: reconstruct the relationship when the surface changes. Do not simply reuse the earlier numerical answer or carry a familiar label into a new situation.
For real-world application, add another layer. A workplace shared task may involve repeated relationships, unequal resources and contributions that are difficult to observe. The token exercise can clarify an incentive conflict, but it does not by itself justify a workplace sanction or predict each employee's behaviour. State which mechanism the analogy illuminates and which important features it leaves outside the model.
CHAPTER 14 OF 14 · Transfer and continue learning
14. Use experiments as a disciplined learning route
A beginner can start with the market-surplus calculation, then explain how price divides surplus without necessarily changing its total. Next, reconstruct the public-goods payoffs and identify the unit of random assignment. Only then move to estimating an effect and interpreting an interval. This sequence connects arithmetic, institutions and evidence.
For a practical study proposal, produce a one-page design note: question, participants, treatment, comparator, assignment unit, outcomes, period, payments, analysis and foreseeable failure points. Ask another reader to identify an alternative explanation. If the note cannot answer that challenge, improve the design before recruiting participants.
Is experimental economics the same as behavioural economics?
No. Experimental economics describes an approach to generating evidence through controlled variation. Behavioural economics concerns theories and findings about decision-making that incorporate psychological and social features. Experiments can test conventional economic predictions as well as behavioural ones. The Behavioural Economics root is the next route for those patterns of choice.
Does a laboratory experiment need to look like everyday life?
It needs to contain the features relevant to its question and make its scope explicit. Simplification can isolate a mechanism. Application to a different setting requires an argument about which features matter and additional evidence where the argument is uncertain. Realistic decoration cannot compensate for an unclear comparison.
Does randomisation prove the mechanism?
Randomisation can identify the effect of an assigned treatment under the design's assumptions. A treatment may change several psychological or practical processes. Establishing which process caused the outcome generally needs additional design or evidence. Measuring a mediator after treatment does not automatically isolate its causal role.
What should a reader inspect in a published result?
Look for the assignment unit, independent sample size, outcome definition, effect magnitude, uncertainty, missing outcomes and deviations from the plan. Check whether the conclusion concerns assignment, participation or a broader policy. Then ask what changed between the experimental environment and the proposed application.
Where should learning continue?
Use Econometrics for statistical interpretation and causal comparisons, Mechanism Design for the relationship between rules and strategic action, and Social Cooperation for shared-resource problems. The economists' professional guide connects evidence to practical trade-offs.
The primary sources linked throughout provide routes into experimental institutions, field settings, randomisation, power, analysis, registration, demand effects, deception and replication. Read them alongside the original worked examples here, keeping a clear distinction between published findings and invented teaching figures. The useful final capability is to explain what an experiment changes, what its data establish, and what would have to be checked before the conclusion travels.