Model Fitting Using the Least Volume Criterion

Abstract

Given data on multiple variables we present a method for fitting a function to the data which, unlike conventional regression, treats all the variables on the same basis i.e. there is no distinction between dependent and independent variables. Moreover, all variables are permitted to have error and we do not assume any information is available regarding the errors. The aim is to generate law-like relationships between variables where the data represent quantities arising in the natural and social sciences. Such relationships are referred to as structural or functional models. The method requires that a (monotonic) relationship exists; thus, in the two variable case we do not allow cases where there is zero correlation. Our fitting criterion is simply the sum of the products of the deviations in each dimension and so corresponds to a volume, or more generally a hyper-volume. One important advantage of this criterion is that the fitted models will always he units (i.e. scale) invariant. We formulate the estimation problem as a fractional programming problem. We demonstrate the method with a numerical example in which we try and uncover the coefficients from a known data-generating model. The data used suffers from multicollinearity and there is preliminary evidence that the least volume method is much more stable against this problem than least squares.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Jul 01, 2001
Accession Number
ADP013739

Entities

People

  • Chris Tofallis

Organizations

  • University of Hertfordshire

Tags

Communities of Interest

  • Air Platforms

DTIC Thesaurus Topics

  • Accuracy
  • Algorithms
  • Coefficients
  • Data Science
  • Data Sets
  • Economics
  • Errors
  • Estimators
  • Information Science
  • Measurement
  • Monotone Functions
  • Monte Carlo Method
  • Normal Distribution
  • Numerical Analysis
  • Standards
  • Statistics
  • Technical Information Centers

Fields of Study

  • Mathematics

Readers

  • Regression Analysis.
  • Theoretical Analysis.

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference
  • AI & ML - Machine Learning Algorithms