Where does discrimination collide?

In perception and learning, discrimination is the ability to tell stimuli apart. In sensory psychophysics it is measured as a difference threshold, or just-noticeable difference: the smallest change in a stimulus that a person can reliably detect, which by Weber's law tends to grow in proportion to the size of the starting stimulus. Gescheider's Psychophysics: The Fundamentals lays out the methods for finding these thresholds and the laws, associated with Weber and Fechner, that relate sensation to stimulus intensity. In conditioning, stimulus discrimination is responding differently to different cues, so that a response follows one signal but not a similar one; it is the complement of /generalization and is built by the discrimination training studied in /conditioning. The referent here is a capacity to separate stimuli, and here finer discrimination is a better outcome, not a harm.[1]

Unit

The smallest difference between two stimuli that can be reliably told apart

Recovery means

The difference threshold: the change in a stimulus detected on some agreed fraction of trials, often half

Measured by

Found by classic psychophysical methods (constant stimuli, limits, adjustment) as a just-noticeable difference, and in learning by different response rates to a trained versus an untrained cue

In statistics and machine learning, discrimination is a model's ability to separate classes. Fisher's The use of multiple measurements in taxonomic problems introduced discriminant analysis, finding the combination of measurements that best pulls apart two or more groups. For a classifier that outputs a score, discrimination is how well those scores rank a true positive above a true negative, which the receiver operating characteristic (ROC) curve displays and the area under it summarises, as Hanley and McNeil described. Discrimination is kept separate from calibration: one asks whether the model separates and ranks cases well, the other whether its predicted probabilities match the rates actually observed. It is a neutral performance property, and more of it means a sharper model.[2][3]

Unit

How well a model's scores separate classes, often the area under the ROC curve from 0.5 to 1

Recovery means

A model with no discrimination scores at chance (area under the curve near 0.5); perfect separation scores 1

Measured by

Quantified by the ROC curve and the area under it (the c-statistic), or by discriminant functions that maximise between-class relative to within-class spread

In economics and social science, discrimination is unequal treatment of people based on group membership rather than on their individual characteristics. Becker's The Economics of Discrimination set out the taste-based account, in which an employer, customer, or coworker holds a preference against a group and acts on it even when doing so is costly. Phelps's The Statistical Theory of Racism and Sexism, alongside work by Arrow, set out statistical discrimination, in which a decision-maker who cannot observe an individual's productivity uses average group traits as a proxy, so members of a group are judged by that average. These are distinct constructs, and researchers test which is at work. The referent is a gap in treatment between groups, and the word names something regarded as a social harm.[4][5]

Unit

A gap in treatment or outcomes between groups not explained by relevant individual differences

Recovery means

Not applicable

Measured by

Studied through wage and hiring gaps adjusted for productivity, and through field experiments such as audit and correspondence studies

In law, discrimination is unequal treatment on a protected ground, such as race or sex, that a statute forbids. United States employment law draws a line between two theories. Disparate treatment is treating a person less favourably because of a protected characteristic; the burden-shifting framework for proving it was set in McDonnell Douglas Corp. v. Green. Disparate impact is a facially neutral practice that falls more heavily on a protected group and is not justified by business necessity; it was recognised in Griggs v. Duke Power Co. The referent is a legal finding, and which theory applies changes what a claimant must show.[6][7]

Unit

A finding that a protected characteristic caused unequal treatment or an unjustified adverse impact

Recovery means

Not applicable

Measured by

Established in court through evidence of intent and comparators (disparate treatment) or statistical disparities weighed against a business-necessity defence (disparate impact)