SOME EFFECTS OF ERRORS OF MEASUREMENT ON LINEAR REGRESSION.
Abstract
When y has a linear regression on X', which is subject to an error of measurement h', there is reason to believe that the regression of y on the fallible x = X+h will not in general satisfy Lindley's conditions for linearity. However, a linear component of this regression can be defined in terms of the low moments of the joint distribution of X' and h', and this component satisfies the usual elementary results given in the literature. Working out the exact regression of y' on x' in a few simple cases suggests (i) that the linear component often dominates even for measurements of only moderate reliability (ii) good approximations to the exact regression can be obtained by quadratics or cubics in x'. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 10, 1970
- Accession Number
- AD0711355
Entities
People
- W. G. Cochran
Organizations
- Harvard University