Where does shrinkage collide?

Wherry's 1931 A new formula for predicting the shrinkage of the coefficient of multiple correlation named the fall in a multiple correlation that appears when a prediction equation fitted on one sample is applied to a fresh sample: the weights that fit the first group partly capture its chance features, so the correlation drops when the equation meets new data. The gap between the correlation in the fitting sample and the correlation on new data is the shrinkage, and Wherry gave a formula to estimate how large it would be before a new sample was even collected. Here the referent is an unwanted loss of apparent predictive accuracy caused by overfitting, not a chosen estimator, a dying cell, or a store's missing stock.[1]

Unit

The difference between a fitted multiple correlation and its value on new data

Recovery means

Not applicable

Measured by

Estimated by an adjustment formula or observed directly by cross-validation on a holdout sample

Stein's 1956 Inadmissibility of the usual estimator for the mean of a multivariate normal distribution showed that when three or more quantities are estimated at once, pulling the sample averages partway toward a common point gives a smaller expected error than using the averages on their own. James and Stein's 1961 Estimation with quadratic loss gave the explicit rule, now called the James-Stein estimator, that carries out this deliberate pull. Here shrinkage is a chosen strategy that trades a little bias for a larger cut in variance, opposite in spirit to the psychometric meaning, where the loss of accuracy is suffered rather than sought.[2][3]

Unit

The amount an estimate is moved toward a target, set by a shrinkage factor

Recovery means

Not applicable

Measured by

Quantified by the shrinkage factor and judged by the reduction in mean squared error against the plain estimator

Kerr's 1971 Shrinkage necrosis: a distinct mode of cellular death described cells that die by condensing: the cell loses water and volume, its contents grow dense, and it breaks into small bodies, unlike cells that swell and burst. Kerr, Wyllie, and Currie's 1972 Apoptosis: a basic biological phenomenon with wide-ranging implications in tissue kinetics renamed this controlled death apoptosis and set cell shrinkage among its defining features. Here shrinkage is the physical loss of volume by a single dying cell, watched under a microscope, and it shares nothing with a statistical estimator, a correlation measured on new data, or stock missing from a shop.[4][5]

Unit

The volume lost by one cell as it dies, a fall from its normal size

Recovery means

The shrunken cell does not recover; it fragments and is cleared by neighbouring cells

Measured by

Seen as reduced size and denser contents under a microscope, and quantified by flow cytometry as a fall in forward scatter

Raman, DeHoratius, and Ton's 2001 Execution: the missing link in retail operations examined the gap between the stock a retailer's records show and the stock actually on the shelf, and treated shrinkage, the goods that are paid for but never sold and cannot be found, as a core failure of store execution. DeHoratius and Raman's 2008 Inventory record inaccuracy: an empirical analysis measured this gap across thousands of stores and traced how theft, damage, and miscounting drive it. Here shrinkage is merchandise recorded in the books but absent from the store, counted in currency or units, and unrelated to a shrinking cell, a pulled estimate, or a fallen correlation.[6][7]

Unit

The value or count of stock the records show but the store cannot produce

Recovery means

Reduced by loss prevention and tighter counting, though rarely brought to zero

Measured by

Calculated as the difference between book inventory and a physical count, reported as a percentage of sales