LB 23-195 - No limits: A call for reporting values below detection along with their uncertainty

Solute concentrations too low to be detected are not uncommon in natural waters. For example, at the Hubbard Brook Experimental Forest in New Hampshire, USA, ammonium and phosphate in streamwater are almost always below detection, and nitrate is below detection in about half of all samples. The method detection limit is the concentration that we have 99% confidence is different from zero. It is difficult to decide how to handle these observations; when calculating values that depend on them, such as ecosystem inputs in precipitation or losses in runoff, values are needed—merely noting that concentrations are below detection is insufficient. We asked questions about handling problematic data in a survey of 140 scientists representing 90 different research sites or networks (Yanai et al. 2017). A surprising number of respondents (15 of 60, or 25%) reported that they replace these values with zeros, which would introduce a bias, as zero is always lower than the true value. A smaller number (9) reported omitting the values, which incurs a bias in the other direction, as omitting low values leaves the others to overestimate the true distribution. Using half the detection limit, which is the current practice at Hubbard Brook, is also common (17 respondents). We contend that it would be more truthful to report the values obtained from the instrument, along with the uncertainty in the values, acknowledging that this will result in some negative values. We calculated the consequences of these decisions for stream export of P from 2013 to 2020 for Hubbard Brook Watershed 3, for which 93% of values were below detection. Using the numbers from the instrument gave 11.1 g P/ha/y. Using zero gave 2.2 g P/ha/y, while omitting values below detection gave 32.8 g P/ha/y. Using half of the detection limit slightly overestimates export calculated from the best estimates, at 15.0 g P/ha/yr. Some scientists advocate using half of the limit of quantification (the concentration where replicates are within 10% of the mean), which gave 90.3 g P/ha/y. Reporting the values obtained from the instrument would be less biased than using half the detection limit, and reporting uncertainty in all values would improve transparency.