Does classical up-and-down usually target a 50% response region?

Yes, under its standard monotone-response setup.

Are assigned levels independent across observations?

No; later levels depend on earlier responses.

Does a response always mean move down?

Only under the stated convention; the meaning of success and direction must be explicit.

AI Glossary / Statistics and probability / Intermediate

Up-and-down design

Definition

An up-and-down design is a sequential experimental design in which the level used for the next observation depends on the preceding response. In the classical binary-response version, a negative response moves the level upward and a positive response moves it downward, assuming response probability increases with level. This concentrates observations near the transition region, commonly around a 50% response probability. Variants can target other probabilities, but their rules and interpretation must be specified.

What it means

Instead of assigning all observations to a fixed grid in advance, the experiment adapts as responses arrive. If a level is too low to produce the response, the next observation moves higher. If the response occurs, the next observation moves lower. Over time the sequence tends to visit levels around the design's balance point. The observations are generated by an adaptive process, so the sequence should not be analyzed as though levels were independently assigned without regard to earlier outcomes.

How it works

Choose a starting level, an ordered set of allowable levels, a step rule, boundary handling, stopping criteria and an analysis method before the experiment. For the classical equal-step rule, move up after response zero and down after response one. At a level with response probability p, the expected direction of the next move is proportional to (1 minus p) minus p, or 1 minus 2p. This balances at p = 0.5 under the idealized local reasoning. Actual finite-sample behavior depends on level spacing, boundaries, starting point and the response curve. Specialized estimators and uncertainty procedures should reflect the design.

Three practical examples

Example 1: A nonclinical perception experiment adjusts a visual stimulus to study the level at which detection becomes likely. Example 2: A materials laboratory varies a test setting based on a binary response to locate a transition region. Example 3: An interface study adapts a task difficulty level after success or failure, with the direction convention adjusted to its response definition. These examples describe experimental design, not universal operating instructions; ethical and domain-specific constraints determine permissible settings.

Think of it this way

Think of a staircase search for a transition. A response tells you to step one way, while its absence tells you to step the other. The staircase spends more time near the boundary between common and uncommon responses. It is a way to choose informative levels, not a guarantee that the last step is the exact threshold.

Worked example

Consider a nonclinical test with integer stimulus levels and the classical rule: move up one after a negative response and down one after a positive response. Start at level four. If successive responses are zero, zero, one, zero, one and one, the administered levels are four, five, six, five, six and five; the next planned level is four. At a hypothetical level where response probability is 0.7, the expected one-step change is 0.3 times one plus 0.7 times minus one, equal to -0.4 levels. At probability 0.5, the expected change is zero. The short observed sequence illustrates allocation only; it is not enough to claim a precise threshold estimate or a confidence interval without a specified estimator and assumptions.

Illustrative case study

In this fictional case study, Larch Displays studies detection of a harmless visual marker. Researchers predefine the allowed brightness levels, step size, boundary behavior and stopping rule. After piloting comprehension and equipment consistency, they use an up-and-down sequence to focus observations near the transition. Their report preserves the full response order and uses an analysis suitable for adaptive allocation. They avoid reporting the final brightness setting as though it were an unbiased, precisely estimated population threshold.

Common misconceptions and mistakes

Common mistakes include confusing the next assigned level with an estimated threshold, ignoring dependence induced by adaptation, and assuming the classical rule targets any desired response percentage. Another mistake is averaging all visited levels without considering the estimator's assumptions or starting effects. Reversals can be informative but do not automatically guarantee adequate precision. A nonmonotone response relationship, strong time trends or inconsistent response definitions can undermine the intended concentration near a transition.

Business use

Businesses and research teams can use the design when observations are costly and a transition level is more relevant than the entire response curve. Define the target probability and select a rule designed for it. Simulate operating characteristics under plausible response curves, including boundaries and stopping conditions. Preserve the sequential data and distinguish exploratory pilots from confirmatory estimates. If the decision needs performance far from the design's balance point, a different allocation or additional measurements may be necessary.

Check your understanding

What level follows the first two negative responses in the example?

Answer: Level six.

What is the next planned level after all six responses?

Answer: Four.

What is the expected step at response probability 0.7?

Answer: -0.4 levels.

What does that drift suggest?

Answer: The classical rule tends to move downward when responses are more common than its balance point.

Related terms

these are related but distinct choices

Learn next

Learn next how different up-and-down rules target different response probabilities and how level spacing affects concentration. Examine simulation-based assessment and estimators tailored to the sequential design. Then compare threshold-focused allocation with designs intended to estimate an entire response curve.

Sources

Source: upndown package authors, Up-and-Down Basics, https://cran.r-project.org/web/packages/upndown/vignettes/upndown_basics.html. This primary implementation guide explains classical allocation, dependence and estimation cautions. The integer-level sequence and expected-step calculation are original and independently checked. The display experiment is fictional and concerns a nonclinical stimulus.

Previous: Unvoiced sound Next: Update Anomalies