Location: Agroecosystems Management ResearchTitle: Concordance correlation for model performance assessment: An example with reference evapotranspiration observations) Author
Submitted to: American Society of Agronomy Meetings
Publication Type: Abstract Only
Publication Acceptance Date: 4/19/2007
Publication Date: 11/8/2007
Citation: Meek, D.W. 2007. Concordance correlation for model performance assessment: An example with reference evapotranspiration observations [CD-ROM]. In: ASA-CSSA-SSSA Annual Meeting Abstracts, November 4-8, 2007, New Orleans, LA. Interpretive Summary:
Technical Abstract: Procedure for assessing model performance in agronomy is often arbitrary and not always helpful. An omnibus analysis, correspondence correlation, is available and widely known and used in many other sciences. An illustrative example is presented here. The analysis assumes the exact relationship observations = predictions is true. An adjusted correlation coefficient (rc) for the exact correspondence model is estimated using adjustments on the well-known product-moment correlation coefficient (r) for a scale shift (v) and a location shift (u). Data are for 50 d selected from a published lysimeter - weather station calibration study. Daily totals of measured reference evapotranspiration (ET0) are compared to estimates from two possible weather observation based models using correspondence correlation analysis. Both models use the same weather data inputs and estimate an hourly total. The daily value for model-1 is the sum of hourly estimates that use a published empirical wind function calibration. The daily value for model-2 is the sum of hourly estimates that use a published iterative aero-dynamical stability routine. Using model-1, r = 0.980, rc = 0.975 (P<0.015), v = 0.908, and u = -0.025. Using model-2, r = 0.982, rc = 0.945 (P<0.037), v = 0.841, and u = -0.215. Hence, although model-2 predictions have less scatter, having a greater r, it is not a better choice because it has both greater u and v values revealing greater systematic under-estimation of observation location (mean bias) and scale (variance). Researchers should therefore consider correspondence correlation analysis for model performance assessment because it provides a simple and sound probability-based omnibus test and useful analytical insight.