Where does accuracy collide?

The proportion of cases that a classification rule or predictive model labels correctly, judged against known outcomes. Yerushalmy's 1947 Statistical Problems in Assessing Methods of Medical Diagnosis gave an early statistical treatment of how well a diagnostic method separates true cases from false ones, showing that a single fraction-correct figure can mislead and introducing separate rates for detecting positives and negatives. Swets's 1988 Measuring the Accuracy of Diagnostic Systems extended this by measuring a system's ability to discriminate independently of the chosen decision threshold. Here accuracy is a correctness rate over discrete decisions, not the closeness of a physical measurement to a true value or the match between a judge's impression and a person's real traits.[1][2]

Unit

A correctness rate over a set of classified cases, from 0 to 1 or as a percentage

Recovery means

Not applicable

Measured by

Computed as the number of correct classifications divided by the total, often reported alongside sensitivity, specificity, or ROC area

The correspondence between one person's judgment of another's characteristics and that person's actual standing on those characteristics, checked against self-reports, behavior, or informant criteria. Cronbach's 1955 Processes Affecting Scores on 'Understanding of Others' and 'Assumed Similarity' showed that a raw match between a judge's ratings and a target can be inflated by response tendencies and by guessing the average person, and separated the components hidden inside such a score. Funder's 1995 On the Accuracy of Personality Judgment reframed the question around when and how judges arrive at correct impressions of real traits. Here accuracy is a property of interpersonal perception, not the correctness rate of a classifier or the error of a physical instrument.[3][4]

Unit

A degree of correspondence between a judge's ratings and a target's criterion standing

Recovery means

Not applicable

Measured by

Estimated by correlating a judge's ratings with the target's self-reports, informant reports, or behavioral criteria

The closeness of a measured value to the true value of the quantity being measured, kept distinct from precision, which is the closeness of repeated measurements to one another. Eisenhart's 1963 Realistic Evaluation of the Precision and Accuracy of Instrument Calibration Systems formalized this split for measurement science, tying accuracy to systematic error, or bias, rather than to random scatter. The international standard ISO 5725 later divided accuracy into trueness, the closeness to a reference value, and precision. Here accuracy describes the error of a single quantitative measurement against a true or reference value, not the fraction of correct categorical predictions or the match between a judge and a person.[5][6]

Unit

The deviation of a measured value from the true or reference value of the quantity

Recovery means

Reduced by calibration against a reference standard that corrects systematic error

Measured by

Assessed by comparing readings to a certified reference or standard and quantifying bias or trueness