“Truncation” is what happens when a sub group is not sampled randomly. Potentially, something in this non-random sample will bias certain variables, or “truncate” the selection, resulting in skewed performance.
Truncation is when a specific portion of the population has incomplete predictive information due to limited selection of that population, such as rejection of a risky population or being turned down by the consumer due to a non-competitive product.
For example, a particular lender may ‘target’ a population with recent bankruptcies, because of their policies, this ‘targeted’ bankrupt population may perform better than the general bankrupt population, resulting in a bankruptcy variable, such as time since last bankruptcy, showing a positive relationship with risk. This is because most of the bad risk bankrupt prospects or applicants were rejected. Thus a model based on this data would score bankruptcies more favorably AND IF the previous policies that were in place to ‘target’ this population were changed it could have disastrous results.
“Truncation” can be a serious problem when there are subgroups that are normally excluded from the “known” population. Often, selected observations will be included that pass some additional criteria.
For example, in a credit situation, sometimes a credit policy will reject applicants with a short term history of severe delinquency; however, selected overrides may be allowed in cases where the applicant has sufficient collateral. This could potentially result in a situation where people with severe DQ have better performance than those that don’t (due to the collateral requirement). Including this pattern in a model, especially if the credit policy is changed, might result in high acceptance rates for people with severe DQ.
Return to FAMQ
Return to FAMQ