A genetic algorithm was implemented to select models to predict fecal, urinary, and total manure nitrogen (N) excretion by lactating dairy cows. Akaike’s Information Criterion was used as the criteria for model selection. Data for model development consisted of 1,047 indirect calorimetry observations on lactating cows collected at the USDA’s Energy and Metabolism laboratory from 1963 to 1995. Two tiers of model classes based on model input requirements were developed resulting in 6 models. Tier 1 selects a single covariate model to estimate N excretion and the Tier 2 selects models from many potential dietary and animal covariates including BW, DIM, CP, NDF, ash content, and proportion of concentrate in the diet. Animal and study were designated as cross-classified random effects and the final selected mixed models were fit using restricted maximum likelihood in R statistical software with the lme4 package. The root mean square prediction error (RMSPE) was used to evaluate the models in 3 ways: (1) K-fold cross validation based on all data, (2) evaluation with the most recent 6 years of data, and (3) evaluation with N balance data collected from literature published from 1996 to 2014. The number of published studies reporting covariates required for tier 2 models was not sufficient for model evaluation. Results listed in Table 1 for tier 1 models show better prediction for total manure N and fecal N compared with urinary N excretion. Tier 2 models had lower RMSPE than Tier 1 models across all forms of excretion. Table 1. Model selection and evaluation reported in the RMSPE (% mean N excretion); parameter estimates with standard errors in parentheses

Model estimates	USDAcross-validation	USDA1990-1995	Literaturedata1995-2014
UN = 12.0(5.80) + 0.333(0.0106) × NI	23.7	12.7	23.8
FN = −18.5(3.59) + 10.1(0.169) × DMI	12.6	18.2	16.2
TN = 20.3(4.72) + 0.654(0.00926) × NI	9.80	8.14	10.8

UN = urine nitrogen, FN = fecal nitrogen, TN = total manure nitrogen, NI = nitrogen intake (g/d).