0 votes
2.2k views
Dear Andreas, Michael, GreenDelta,

Congratulations on the new version of openLCA 1.6 and also updating the Monte Carlo feature. I can now confirm that the negative results previously encountered with openLCA 1.3/1.4/1.5 are no longer there.

I do have twp separate questions and they pertain to the sampling approach used in openLCA for its MC function:

(1) Can the sampling approach in openLCA be summarized as "partially independent" sampling as is clarified in Suh and Qin 2017? (https://link.springer.com/article/10.1007/s11367-017-1287-x)

(2) Is it possible to isolate sampling only for the foreground processes? (i.e. turning off all upstream sampling)

All the best,

Michael
in openLCA by 1 flag

1 Answer

0 votes
by (113k points)
 
Best answer
Hi Michael,

thank you for the congrats!

To your questions:
-> 1) We perform the simulation in openLCA in a way that would be classified as "fully dependent" in the article you mention: all uncertain data is drawn at the same time, and then a calculation is started. If a process appears in several "branches" of a supply chain several times, uncertain data is drawn for this process only once and then used in the calculation.
-> 2)"is it possible to isolate sampling only for the foreground processes? (i.e. turning off all upstream sampling)": unfortunately not but I agree this would be a nice feature.

Best wishes,
Andreas
by
Hi Andreas,

Thank you for the prompt reply and clarification. This is very good to know about the dependent sampling.

Perhaps in the future, we can all look forward to an added feature with isolated sampling (e.g. foreground versus background sampling).

One additional question, are the means and standard deviations always estimated under the assumptions of a specific distribution (i.e. does openLCA assume that all impacts are normally or log-normally distributed)?

All the best,

Michael
by (113k points)
Hi Michael,
the distribution is used as specified by the database or user (of course, I would say - taking always the same distribution would be quite plain).
And I agree to your suggestion, to look for possibilities to implement background vs. foreground sampling.
Best wishes
Andreas
by
Dear Andreas,

My apologies, I didn`t make clear my question. Defining the inventory distributions is clear for me. But I am wondering how openLCA calculates the mean and standard deviation for the resulting impacts (i.e. after the impacts are calculated for each Monte Carlo simulation, is the mean and standard deviation calculated assuming the simulations are normal distributed, or are the mean and standard deviations calculated assuming log-normally distributed impact results)?

Best,

Michael
by (113k points)
Ah - well the mean of the MC simulation result is calculated as arithmetical mean of the various items in the result. There is no assumption about distribution needed (which is one of the strengths of the Monte Carlo simulation). That was your question?
by
Dear Andreas,

Thanks for the reply.

So, the final mean and standard deviation for each impact category are reported as arithmetic means and standard deviations, correct?

This is problematic for skewed data, no? For instance, if the histogram of 10,000 MC outputs follows a log-normal best fit, it would be incorrect to report the impacts with an arithmetic mean and standard deviation.

Best,

Michael
by (113k points)
well,
"This is problematic for skewed data"
I agree, this is a common issue in data analysis, the mean is sensitive to outliers (or, asymetric distributions more generally) and therefore it makes sense to also calculate the median, which we do. It is also possible in openLCA to export all simulation data and perform more analyses, e.g. related to an assumed distribution, with this data. I do not think, though, that there is a right or wrong, correct or incorrect, here, as often in data analysis, but of course the arithmetical mean is not perfect, as other indicators too.
Cheers
Andreas
by
Dear Andreas,

Correct, I have already exported all the data and made the appropriate estimations of the log-transformed data.

For me, I was just curious whether openLCA might include an option to automatically report log-transformed statistics, given that inventory data and thus results may commonly follow log-normal distributions? No need to respond to that, just a suggested-feature to add from a dedicated user :)

All the best,

Michael
by (120 points)
Hello Andreas,

You mentioned, “all uncertain data is drawn at the same time, and then a calculation is started”. So the dependency between unit processes to keep mass/energy balance is not considered. Right?

Thank you,
Shiva
...