Review questions
Integrated summary: Foundations of Behavioral Economics and Evidence
This subject starts with a simple idea: behavioral economics does not ask us to discard economic models. It asks when additional psychological concepts help explain choice and can be tested against evidence. A benchmark model organizes alternatives, constraints, and outcomes; we then ask whether the data suggest a role for attention, representation, intertemporal trade-offs, or social expectations. These concepts must remain testable. Departure from a normative benchmark does not diagnose a person as “irrational,” and a bias label does not become an explanation merely because it fits an outcome after the fact.
That is why we separated three questions. A descriptive question asks what happened: for example, 100 of 250 people who started a form completed it, or 40%. A causal question asks why it happened and what effect a specific change had; the 40% rate alone cannot answer that. A normative question asks what should be chosen under an explicit criterion such as cost, benefit, and rights. If we jump from the rate to saying users are careless, that a specific interface caused withdrawal, or that 40% is unacceptable, we have mixed three different types of inference.
We then moved to families of behavioral hypotheses. Information can exist without entering attention clearly at the moment of choice; judgments can also use shortcuts such as availability, representativeness, or anchoring in specific tasks. A time conflict can arise when someone plans to bear a cost later but postpones when the cost becomes immediate while information is unchanged, a pattern consistent with present bias. Fairness, reciprocity, or expectations about others can also enter choice. Each family, however, needs a prediction that distinguishes it from alternatives. A successful reminder does not prove limited attention if available time changed too, and rejection of an unequal split does not prove that fairness was an individual’s motive.
Good behavioral evidence matches the strength of inference to the strength of design. Observational data help describe patterns and locate problems, but may mix the cause of interest with other factors. Laboratory work can isolate a mechanism more cleanly; field work can examine behavior in an actual process. Neither is universally “better.” For causal inference we seek a comparison that represents, as credibly as possible, what would have happened without treatment. Random assignment, when appropriate and well implemented, helps build that comparison, but it does not guarantee that the result generalizes to every population, culture, or time.
In the numerical example, 220 of 400 treatment users and 180 of 400 comparison users complete. The rates are 55% and 45%, a difference of 10 percentage points. If we ask for the relative increase over the comparison rate, it is 10 ÷ 45 ≈ 22.2%. That 22.2% is not “22.2 points.” The arithmetic also does not establish statistical significance by itself. Likewise, if completion rises from 40% to 50% after a page redesign, the difference is 10 points and the relative increase is 25%; but if traffic source changes at the same time, a simple before/after comparison cannot isolate the redesign’s effect.
Replication and generalization require the same discipline. Many Labs 2 produced a statistically significant same-direction effect in 15 of 28 replication attempts within that project. A replication project of economics laboratory experiments found 11 of 18 results significant in the same direction, with replicated effects smaller on average. These are results for selected study sets and specified procedures, not universal rates for the “truth” of a science or the probability that any new experiment will succeed. The lesson is that confidence accumulates across evidence, and effect size, context, and design belong in the interpretation.
Finally, we linked evidence to responsible diagnosis. Behavioral diagnosis describes a behavior, not a person: who does what, where and when, what baseline applies, and what outcome is measured? In the fictional funnel, 780 users reach the document step and 420 begin uploading. Thus 360 do not begin, and 360 ÷ 780 ≈ 46.2%. This locates a gap but does not identify its cause. Requirements may be unclear, the document may be unavailable, or users may change their mind after reading the terms. Each hypothesis makes a different prediction and suggests a different test.
A stronger hypothesis specifies population, context, change, outcome, comparison, mechanism, and what would reduce confidence in the explanation. For example: among users reaching the document step who already have the required document, showing a clear requirements list earlier may increase upload starts relative to the current presentation if ambiguity is the mechanism; if comprehension improves but upload starts do not change, confidence in that hypothesis falls. Behavioral economics then becomes a disciplined learning process rather than the advance selection of a behavioral “trick.”
Ethics enters from the start. Use the minimum data needed, consider privacy, consent, fairness, and transparency, measure unintended effects, and do not hide material information merely to increase conversion. The ability to influence behavior is not a license to manipulate. Before moving on, keep this workflow: describe behavior precisely; separate description from causality and normative judgment; write competing hypotheses; specify evidence that distinguishes them; choose a design that matches the question; and review limits of inference and ethics before scaling.
Review questions
Test your understanding with these questions. Marking and saved results are available with a free account.
Create a free account to mark your answers and save your progress.