Dear community,
I've been reading a lot about dealing with omics data that lies outside the limit of quantification - there are a bunch of different recommendations on how to approach this. One paper drew the ULOQ - is that a sound idea? does anyone have comments on this /further suggestions?
I am looking forward to your responses.
Jochen Wilhelm
great, thank you for your input!as you suggested a paper on Tobit analysis - would you say that the R package mice might be suited for this?
If there aren't too many censored values on the low end (< LLOQ), what precise values these take shouldn't matter much in the analysis. (Or, we shouldn't do an analysis where the precise values of these observations matter much). If there are many < LLOQ values, then you might consider a different analysis.
Just thinking about it, I'm not sure what to do with values censored at the high end (> HLOQ). I would be wary of replacing them some kind of random distribution with no upper limit, since it's possible to get a really high value by chance. I have done analyses where I replaced > HLOQ with just HLOQ; because there weren't that many values, it didn't really affect the overall conclusions.
Section 4.7, VALUES BELOW DETECTION LIMITS, of the following (USEPA, 2000) has some simple advice about values less than the detection limit.
They say, if there are < 15 % of values below the detection limit, "Replace nondetects with DL/2, DL, or a very small number."
They have more involved methods when Non-detects are 15%–50% of observations.
https://www.epa.gov/sites/default/files/2015-06/documents/g9-final.pdf
For observations < LLOQ, maybe random values from a uniform distribution would make more sense than from a normal distribution.
Dear Sal Mangiafico - thank you very much for your detailed answer! This has helped me a lot