Pre-calculated LCIs with uncertainties revisited

Abstract

Heijungs et al. (Int J Life Cycle Assess 22(3):461, 2017) argued that (1) dependent sampling is required when analyzing uncertainties using Monte Carlo simulation in a comparative life cycle assessment (LCA) and that (2) using pre-calculated geometric standard deviations (GSDs) will lead to a “large overestimation” of the uncertainty of the results in a comparative LCA. While we agree with Heijungs and colleagues on the former from a theoretical standpoint, we find that the latter needs to be substantiated empirically. With this article, we tested the hypothesis that the use of pre-calculated life cycle inventories (LCIs) leads to an overestimation of uncertainty. In addition, we clarified a number of relevant issues and misconceptions in and around the independent sampling. We compared the GSDs of LCIs for 100 randomly selected Ecoinvent processes that were derived (1) using pre-calculated GSD values (partially independent sampling PIS) and (2) without using pre-calculated values (fully dependent sampling FDS). We tested the null hypothesis, E(GSD FDS /GSD PIS ) < 1, using a t test. The difference in the GSDs from fully dependent sampling and partially independent sampling (using pre-calculated LCIs) was statistically significant, though the average ratio GSD FDS /GSD PIS was very close to 1. Furthermore, the average ratio was larger than 1 with the confidence interval ranging from 1.0681 to 1.0696 at the 95% confidence level. Therefore, the null hypothesis was rejected. According to our empirical analyses, the use of pre-calculated LCIs and associated GSDs leads to a slight underestimation rather than an overestimation—not to mention “large overestimation”—of uncertainties in the results. GSDs in the results derived from pre-calculated LCIs and associated GSDs are very close to those derived from fully dependent sampling, and therefore, in practice, pre-calculated LCIs can be used as a reasonable proxy in understanding the uncertainties and variabilities of both non-comparative and comparative LCAs. Given the high uncertainty in the uncertainty analysis itself, the minor difference between GSDs from pre-calculated values and fully dependent sampling should not prevent LCA practitioners from using pre-calculated LCIs and corresponding GSDs for uncertainty analysis when no better option is reasonably practicable.