Positive Matrix Factorization for dummies

The theory behind Positive Matrix Factorization (PMF) -a type of factor analysis- explained using soda as an analogy

[CPO_]

Thank you so much for this concise explanation of PMF (as well as the touch on the other factor analysis methods). I wish you could cover the estimation of uncertainties of analyte measurements, if it's an essential input to PMF analysis - OK, I just saw your explanation of uncertainty estimation in the comments below...😅 THANK YOU!

[wildvineyard]

Uncertainty matrix can be created using: a) relative standard deviations of surrogate recoveries, b) relative percent differences of duplicate measurements, and c) wild ass guesses based on info in the analytical method. A good place to start is to assume a base uncertainty (for detected stuff) of between 10% and 35% and then apply three times this uncertainty to anything that was not detected.

[wildvineyard]

HI all, I frequently get asked about how to construct your uncertainty matrix. Here is my advice: calculating uncertainty is tough. We use the RSD of surrogate recoveries, but if you don't have those, you can use the Relative Percent Difference of duplicates. Often the RPD ranges from tiny to >100%, so I usually scale it so that the uncertainties range from about 10% to about 35% but the ranking of the uncertainties among the analytes is the same. I have also used information on reproducibility that we got from the published method.

If all else fails, you can punt and use a sliding scale, starting with 10% for the things that you feel in your gut are certain (probably including things that are well above the LOD) and use 35% for the less certain things (including stuff near the LOD). For example, I would give a high uncertainty to things like nitrite since it is so reactive and hard to measure, and low uncertainty to things like TOC, which is pretty easy to measure. Keep in mind that PMF 5.0 takes the uncertainty in the same units as the input concentrations, so you have to multiply these percentages by the measured concentration. And you always give a much higher unc to NDs. I usually use three times the regular uncertainty for that analyte and apply it to NDs (multiplied by whatever value I'm using as a proxy for the ND, usually half the LOD).

You may have to try a few unc matrixes until you find one that works. And PMF 5.0 has that field under Base where you can add extra modeling uncertainty. Try it with 0% extra uncertainty and then keep increasing the extra uncertainty by 5% at a time until you get something that works.

[lyesMCA]

can you in video explain how to calculate specie specific uncertainty in PMF in detail. Thank you in advance.

[MdMamun-yewo]

How can I do this model (source apportionment of Pollution by statistical techniques, multiple linear regression (MLR) and absolute principal component scores (APCS))? Could you please help me to do this one?

[zozoelsaadani]

how to create uncertainly file?
i need help

[eda.]

preamp....pop filter....more better butter. Easier than the above equations to understand.