Does the use of pre-calculated uncertainty values change the conclusions of comparative life cycle assessments? – An empirical analysis

In life cycle assessment (LCA), performing Monte Carlo simulation (MCS) using fully dependent sampling typically involves repeated inversion of a technology matrix for a sufficiently large number of times. As the dimension of technology matrices for life cycle inventory (LCI) databases grows, MCS using fully dependent sampling is becoming a computational challenge. In our previous work, we pre-calculated the distribution functions of the entire LCI flows in the ecoinvent ver. 3.1 database to help reduce the computation time of running fully dependent sampling by individual LCA practitioners. However, it remains as a question whether the additional errors due to the use of pre-calculated uncertainty values are large enough to alter the conclusion of a comparative study, and, if so, what is the odds of such cases. In this study, we empirically tested the probability of altering the conclusion of a comparative LCA due to the use of pre-calculated uncertainty values. We sampled 10,000 random pairs of elementary flows of ecoinvent LCIs (ai and bi) and ran MCSs (1) using pre-calculated uncertainty values and (2) using fully dependent sampling. We analyzed the distribution of the differences between ai and bi (i.e., ai−bi) of each run, and quantified the probability of reversing (e.g., ai > bi became ai bi became ai ≈ bi). In order to better replicate the situation under a comparative LCA setting, we also sampled 10,000 random pairs of elementary flows from the processes that produce electricity, and repeated the same procedure. The results show that no LCIs derived using pre-calculated uncertainty values constitute large enough differences from those using fully dependent sampling to reverse the conclusion. However, in 5.3% of the cases, the conclusion from one approach is moderated under the other approach or vice versa. When elementary flow pairs are sampled only from the electricity-producing processes, the probability of moderating the conclusions increases to 10.5%, while that of reversing the conclusions remains nil. As the number of unit processes in LCI databases increases, running full MCSs in a PC-environment will continue to be a challenge, which may lead some LCA practitioners to avoid uncertainty analysis altogether. Our results indicate that pre-calculated distributions for LCIs can be used as a proxy for comparative LCA studies in the absence of adequate computational resources for full MCS. Depending on the goal and scope of the study, LCA practitioners should consider using pre-calculated distributions if the benefits of doing so outweighs the associated risks of altering the conclusion.