A simulation study of nonparametric total deviation index as a measure of agreement based on quantile regression.

Published

Journal Article

Total deviation index (TDI) captures a prespecified quantile of the absolute deviation of paired observations from raters, observers, methods, assays, instruments, etc. We compare the performance of TDI using nonparametric quantile regression to the TDI assuming normality (Lin, 2000). This simulation study considers three distributions: normal, Poisson, and uniform at quantile levels of 0.8 and 0.9 for cases with and without contamination. Study endpoints include the bias of TDI estimates (compared with their respective theoretical values), standard error of TDI estimates (compared with their true simulated standard errors), and test size (compared with 0.05), and power. Nonparametric TDI using quantile regression, although it slightly underestimates and delivers slightly less power for data without contamination, works satisfactorily under all simulated cases even for moderate (say, ≥40) sample sizes. The performance of the TDI based on a quantile of 0.8 is in general superior to that of 0.9. The performances of nonparametric and parametric TDI methods are compared with a real data example. Nonparametric TDI can be very useful when the underlying distribution on the difference is not normal, especially when it has a heavy tail.

Full Text

Duke Authors

Cited Authors

  • Lin, L; Pan, Y; Hedayat, AS; Barnhart, HX; Haber, M

Published Date

  • 2016

Published In

Volume / Issue

  • 26 / 5

Start / End Page

  • 937 - 950

PubMed ID

  • 26391352

Pubmed Central ID

  • 26391352

Electronic International Standard Serial Number (EISSN)

  • 1520-5711

Digital Object Identifier (DOI)

  • 10.1080/10543406.2015.1094812

Language

  • eng

Conference Location

  • England