Berkson's paradox, also known as Berkson's bias or collider bias, is a counterintuitive phenomenon that arises in conditional probability when there is an ascertainment bias in a study's design. This can lead to spurious statistical associations, where two independent variables appear to be negatively correlated in a selected population sample, despite being uncorrelated or positively correlated in the general population.
For instance, a dating pool may present an observed negative correlation between 'niceness' and 'handsomeness' among men that one person dates. Someone might think that nicer men tend to be less handsome, even though these traits are actually uncorrelated in the wider male population.
To overcome Berkson's paradox, ensure that the sample population is representative of the general population. Avoid conditional sampling that may create intrinsic biases in the observed data.