Model robustness for simultaneous confidence bands

Research output: Contribution to journalArticlepeer-review

4 Scopus citations


Confidence bands for the simple linear model are examined to assess their degrees of robustness to departures from the model. All calculations are made under an interval constraint on the range of interest for the predictor variable. The true model is taken to be a quadratic polynomial, and departure from the linear case is considered in terms of increasing magnitude of a curvature parameter. Proposed measures of robustness include the actual coverage probability under the true model and a measure of percentage coverage over the predictor variable axis when the band fails to cover the true quadratic model. The different band functions considered include hyperbolic bands constructed from Scheffd’s 5 method, linear-segmented bands, and fixed-width (minimax) and minimax-regret bands. In terms of preserving coverage probability under quadratic mis-specification, the fixed-width and linear-segment bands perform best, the former being preferred when the constraint interval on the predictor variable is small. When coverage is lost over some portion of the constraint interval, the fixed-width bands are also shown to preserve the greatest percentage of covered (predictor) axis values. The favorable performance of the fixed-width bands may be due to their generally wider, more rigid shape, relative to more curvilinear competitors. This may allow the bands to retain more extreme quadratic misspecifications than other bands.

Original languageEnglish (US)
Pages (from-to)879-885
Number of pages7
JournalJournal of the American Statistical Association
Issue number399
StatePublished - Sep 1987


  • Coverage probability
  • Mean axis coverage
  • Quadratic regression
  • Simple linear regression

ASJC Scopus subject areas

  • Statistics and Probability
  • Statistics, Probability and Uncertainty


Dive into the research topics of 'Model robustness for simultaneous confidence bands'. Together they form a unique fingerprint.

Cite this