My Account: Log In | Join | Renew
Search
Author
Title
Vol.
Issue
Year
1st Page

Abstract

 

This article in AJ

  1. Vol. 101 No. 4, p. 1012-1018
     
    Received: Nov 12, 2008
    Published: July, 2009


    * Corresponding author(s): Dave.Meek@ars.usda.gov
 View
 Download
 Alerts
 Permissions
 Share

doi:10.2134/agronj2008.0180x

Concordance Correlation for Model Performance Assessment: An Example with Reference Evapotranspiration Observations

  1. David W. Meek *a,
  2. Terry A. Howellb and
  3. Claude J. Phenec
  1. a USDA-ARS, National Soil Tilth Lab., 2110 University Blvd., Ames, IA 50011
    b USDA-ARS Conservation and Production Research Lab., P.O. Drawer 10, Bushland, TX 79012
    c USDA-ARS, SDI+, 13089 Wiregrass Lane, Clovis, CA 93619

Abstract

The assessment procedures for agronomic model performance are often arbitrary and unhelpful. An omnibus analysis, the concordance correlation coefficient (r c), is widely used in many other sciences. This work illustrates model assessment with two r c measures accompanied with a mean-difference (MD) plot and a distribution comparison. Each r c is an adjusted value of the usual Pearson correlation coefficient, r, assuming the exact relationship observations = predictions The adjustments use a scale shift, u, and a location shift, v Both of these measures also can indicate the similarity of the two variables' distributions; however, a formal test, the Kolmorgov-Smirnov D statistic, is used to statistically compare the distributions. Daily evapotranspiration data (ET0) from a published study are compared with estimates from two possible weather observation based models. Although the first model has slightly lower r than the second (0.980 vs. 0.982), its predictions reasonably agree with observations by having comparatively small location and scale shifts [(u = 0.025, v = 1.10, D = 0.12 (p ≤ 0.86)] and, consequently, a higher r c (0.975 vs. 0.946). Results for the second model are comparatively unacceptable having larger scale and location shifts (u = 0.215, v = 1.19, D = 0.28 [p ≤ 0.04]) with the bias ≠ 0 (p ≤ 0.05) as clearly shown in the associated MD plot. Researchers should consider using r c with an MD plot and distribution comparison in their model assessment toolkit because, together, they can provide a simple and sound probability based omnibus test as well as add useful insight.

  Please view the pdf by using the Full Text (PDF) link under 'View' to the left.

Copyright © 2009. American Society of AgronomyCopyright © 2009 by the American Society of Agronomy