Empirically based uncertainty factors for the pedigree matrix in ecoinvent

Ecoinvent applies a method for estimation of default standard deviations for flow data from characteristics of these flows and the respective processes that are turned into uncertainty factors in a pedigree matrix, starting from qualitative assessments. The uncertainty factors are aggregated to the standard deviation. This approach allows calculating uncertainties for all flows in the ecoinvent database. In ecoinvent 2 the uncertainty factors were provided based on expert judgment, without (documented) empirical foundation. This paper presents (1) a procedure to obtain an empirical foundation for the uncertainty factors that are used in the pedigree approach and (2) a proposal for new uncertainty factors, received by applying the developed procedure. Both the factors and the procedure are a result of a first phase of an ecoinvent project to refine the pedigree matrix approach. A separate paper in the same edition, also the result of the aforementioned project, deals with extending the developed approach to other probability distributions than lognormal (Muller et al.). Uncertainty is defined here simply as geometric standard deviation (GSD) of intermediate and elementary exchanges at the unit process level. This fits to the lognormal probability distribution that is assumed as default in ecoinvent 2, and helps to overcome scaling effects in the analysed data. In order to provide the required empirical basis, a broad portfolio of data sources is analysed; it is especially important to consider sources outside of the ecoinvent database to avoid circular reasoning. The ecoinvent pedigree matrix from version 2 is taken as a starting point, skipping the indicator sample size since it will not be used in ecoinvent 3. This leads to a pedigree matrix with five data quality indicators, each having five score values. The analysis is conducted as follows: for each matrix indicator and for each data source, indicator scores are set in relation to data sets, building groups of data sets that represent the different data quality indicator scores in the pedigree matrix. The uncertainty in each of the groups is calculated. The uncertainty obtained for the group with the ideal indicator score is set as a reference, and uncertainties for the other groups are set in relation to this reference uncertainty. The obtained ratio will be different from 1, it represents the unexplained uncertainty, additional uncertainty due to a lower data quality, and can be directly used as uncertainty factor candidates. The developed approach was able to derive empirically based uncertainty factor candidates for the pedigree matrix in ecoinvent. Uncertainty factors were obtained for all data quality indicators and for almost all indicator scores in the matrix. The factors are the result of the first analysis of several data sources, further analyses and discussions should be used to strengthen their empirical basis. As a consequence, the provided uncertainty factors can change in future. Finally, a few of the qualitative score descriptions in the pedigree matrix left room for interpretation, making their application not ambiguous. An empirical foundation for the uncertainty factors in the pedigree matrix overcomes one main argument against their use, which in turn strengthens the whole pedigree approach for quantitative uncertainty assessment in ecoinvent. This paper provides an approach to obtain an empirical basis for the uncertainty factors, and it provides also empirically based uncertainty factors, for indicator scores in the pedigree matrix. Basic uncertainty factors are not provided, it is recommended to use the factors from ecoinvent 2 for the time being. In the developed procedure, using GSD as the uncertainty measure is essential to overcome scaling effects; it should therefore also be used if the analysed data do not follow a lognormal distribution. As a consequence, uncertainty factors obtained as GSD ratios need to be translated to range estimators relevant for these other distributions. Formulas for this step are provided in a separate paper (Muller et al.). The work presented in this paper could be the starting point for a much broader study to provide a better basis for input uncertainty in LCA, not only in ecoinvent.