Statistics Glossary → Truncation
Truncationconcept
A situation in which units beyond a certain limit never enter the sample, so neither their values nor their existence is observed.
Truncation and censoring are often confused. With censoring the unit is in the sample but its value is known only to lie beyond a limit (e.g. 'lived at least 10 years'); with truncation, units beyond the limit are not recorded at all. For example, in data from an aid programme that only households below a certain income can apply to, higher-income households are absent from the data set. Ordinary regression is then biased, because the sample was selected on the value of the outcome; slopes are usually pulled toward zero.
Truncated regression assumes the outcome follows a truncated distribution (e.g. truncated normal), rescales the likelihood to the observable region only, and estimates by maximum likelihood. The results are sensitive to the distributional assumption. In count data, situations where zeros are never observed are handled with zero-truncated models; the count part of hurdle models is also a zero-truncated distribution. In Bayesian analysis, a truncated distribution means using only one half of the prior for a directional hypothesis. Truncation is a special case of the more general selection bias that a selection mechanism introduces into the sample.
An everyday example
A fisheries researcher measures the length of fish in a net, but the mesh lets fish shorter than 20 cm escape. Small fish are entirely missing from the data; the measured fish average 31 cm, while the true average in the lake is lower. Truncated regression, knowing the data are 'cut off' below 20 cm, tries to estimate the lake's distribution and the relationship between length and age without this bias.
In MerQur
Available in sample selection and truncated regression as the 'Lower truncation limit' and 'Upper truncation limit' options (active only in truncated regression); observations outside the limits are removed and reported, and the model is estimated by maximum likelihood in a sample that never sees observations outside the limits. Tobit regression handles censored data; the hurdle model of Poisson / negative binomial regression uses a zero-truncated count distribution, and the Bayesian t-test and Bayesian correlation give truncated posterior summaries for directional hypotheses.
Analyses where this concept appears
Poisson / Negative Binomial Regression (Count Regression)Robust Regression (M-estimator)Tobit Regression (Censored Regression)Bayesian t-Test (JZS Bayes Factor)Bayesian CorrelationSample Selection and Truncated Regression (Heckman)
Related concepts
Sample selection bias (Heckman)Censoring (survival analysis)