Sampling error in US field crop unit process data for life cycle assessment
- First Online:
- Cite this article as:
- Cooper, J.S., Kahn, E. & Ebel, R. Int J Life Cycle Assess (2013) 18: 185. doi:10.1007/s11367-012-0454-3
- 415 Downloads
The research presented here was motivated by an interest in understanding the magnitude of sampling error in crop production unit process data developed for life cycle assessments (LCAs) of food, biofuel, and bioproduct production. More broadly, uncertainty data are placed within the context of conclusive interpretations of comparative bioproduct LCA results.
Data from the US Department of Agriculture's Agricultural Resource Management Survey were parameterized for 466 crop–state–year combinations, using 146 variables representing the previous crop, tillage and seed operations, irrigation, and applications of synthetic fertilizer, lime, nitrogen inhibitor, organic fertilizer, and pesticides. Data are described by Student's t distributions representing sampling error through the relative standard error (RSE) and are organized by the magnitude of the RSE by data point. Also, instances in which the bounds of the 95 % confidence intervals are less than zero or exceed actual limits are identified.
Results and discussion
Although the vast majority of the data have a RSE less than 100 %, values range from 0 to 1,600 %. The least precision was found in data collected between 2001 and 2002, in the production of corn and soybeans and in synthetic and pesticide applications and irrigation data. The highest precision was seen in the production of durum wheat, rice, oats, and peanuts and in data representing previous crops and till and seed technology use. Additionally, upwards of 20 % of the unit process, data had 95 % confidence intervals that are less than or exceed actual limits, such as an estimation of a negative area or a portion exceeding a total area, as a consequence of using a jackknife on subsets of data for which the weights are not calibrated explicitly and a low presence of certain practices.
High RSE values arise from the RSE representing a biased distribution, a jackknife estimate being nearly zero, or error propagation using low-precision data. As error propagates to the final unit process data, care is required when interpreting an inventory, e.g., Monte Carlo simulation should only be sampled within the appropriate bounds. At high levels of sampling error such as those described here, comparisons of LCA bioproduct results must be made with caution and must be tested to ensure mean values are different to a desired level of significance.