Why statistical testing and confidence intervals should not be used in comparative life cycle assessments based on Monte
- PDF / 492,938 Bytes
- 5 Pages / 595.276 x 790.866 pts Page_size
- 43 Downloads / 175 Views
COMMENTARY AND DISCUSSION ARTICLE
Why statistical testing and confidence intervals should not be used in comparative life cycle assessments based on Monte Carlo simulations Claudia von Brömssen 1
&
Elin Röös 2
Received: 24 September 2020 / Accepted: 28 September 2020 # The Author(s) 2020
Abstract In the last years, it has been suggested to use statistical inferential methods, such as hypothesis testing or confidence intervals, to compare different products, services, or systems within comparative life cycle assessments based on Monte Carlo simulation results. However, the use of statistical inferential methods in such settings is fundamentally incorrect and should not be continued. In this article, we explain why and look closer at some related topics. Keywords Statistical hypothesis testing . Inferential statistics . Monte Carlo simulations . Uncertainty . Paired simulations
1 Introduction Descriptive statistics, e.g., mean values, medians, variances, or percentiles, can be computed to summarize basically any dataset available. Some caution needs to be taken when choosing the most informative metric for a dataset, depending on dataset characteristics and data collection processes, e.g., while mean values are a good measure for symmetric distributions, they are easily affected by single outliers or when the distribution is skewed. In such cases, a median could be a better representation of the data. Commonly, due to data collection being expensive and timeconsuming, data for the complete population cannot be collected. Imagine, for example, that we want to know the average pesticide use for different crops for a specific year and region and that this information only sits with each individual farmer. It would be too expensive to collect the data from every farmer; instead we would
Communicated by Matthias Finkbeiner * Claudia von Brömssen [email protected] 1
Division of Applied Statistics and Mathematics, Department of Energy and Technology, Swedish University of Agricultural Sciences, P.O. Box 7032, 750 07 Uppsala, Sweden
2
Department of Energy and Technology, Division of Agricultural Engineering, Swedish University of Agricultural Sciences, P.O. Box 7032, 750 07 Uppsala, Sweden
randomly select a sample (a number of observations) of farmers and collect the data from these farmers only. We can then calculate, e.g., the sample mean of pesticide use for a specific crop and then use this information together with the uncertainty of the mean (the standard error) to draw conclusions on the true mean for the whole population (of all farmers growing this crop). This is called inferential statistics. Another typical example would be to investigate if the difference between mean pesticide use for different crops is significant, i.e., if it is reasonable to assume that the two population means of pesticide use are, in fact, different. Inferential statistics is, thus, the area of statistics where descriptive measures are combined with probability theory to make conclusions from a data sample to an underly
Data Loading...
