Where does selection collide?

The differential survival and reproduction of heritable variants in a population, so that traits leaving more descendants become more common over generations. Darwin set out the mechanism in On the Origin of Species in 1859: where individuals vary, where variation is inherited, and where more are born than can survive, the variants that do better at surviving and breeding are selected, and the population changes without anyone choosing anything. The referent is a population across generations, and the unit that is selected is the heritable variant. This is a process that builds adaptation over time, not a distortion in a dataset and not a hiring decision; the word names an outcome of blind differential reproduction, not an act of choosing.[1]

Unit

A heritable variant in a breeding population

Recovery means

Not applicable

Measured by

Change in trait or allele frequency across generations, or fitness differences between variants

A bias in an estimate that arises because the units observed are not a random draw from the population, but were selected in by a process related to the outcome under study. Heckman formalised the sample selection case in 1979, showing that when the chance of being observed depends on the same factors as the dependent variable, an ordinary regression is misspecified, and he gave a two-step correction. The related collider case, described by Berkson in 1946, is where conditioning on a common effect induces a spurious association among its causes. Here selection names a flaw in the data-generating process, a threat to a coefficient, not a force acting on organisms and not a personnel choice; the thing selected is the observation, and the damage is to inference.[2][3]

Unit

The set of observations that enter an estimate

Recovery means

The two-step correction and related methods aim to recover the estimate the selected sample would otherwise hide

Measured by

Bias in a coefficient under a non-random selection mechanism, and its correction

The practice of choosing among job applicants using predictors whose relationship to later performance has been validated. Schmidt and Hunter summarised eighty-five years of evidence in 1998, ranking selection methods by their validity for predicting job performance and finding general mental ability tests, work samples, and structured interviews among the strongest. The referent is an applied decision an organisation makes about individual people, and the unit is the applicant. Unlike natural selection this is a deliberate human choice guided by measurement, and unlike sample selection it is the intended goal rather than a bias to be removed; the same word covers a blind process, a statistical fault, and a designed decision.[4]

Unit

An applicant being chosen or rejected

Recovery means

Not applicable

Measured by

Predictive validity of a method against later job performance

The mind's early filtering of incoming information, so that some inputs are passed on for processing while others are held back. Broadbent proposed selective attention in Perception and Communication in 1958 as a filter model: faced with more sensory input than it can handle, the system selects one channel by its physical features and attenuates the rest, which explains why an unattended message is largely lost. The referent is one perceiver's momentary processing, and the unit selected is a stream of stimuli, not a variant, an observation, or an applicant. The word here names an internal gating of perception, a mechanism of limited capacity, unrelated to reproduction, sampling, or hiring.[5]

Unit

A channel of sensory input within one perceiver

Recovery means

Not applicable

Measured by

Performance on selective- and divided-attention tasks such as dichotic listening