Table 6 Consider Raising Method Reporting Limits (MRLs)

Approach (see Additional file 4 for example of this approach with real data):

1A. For chemicals not detected in blanks, the MRL is equal to the laboratory reporting limit.

1B. For each chemical detected in blanks, if there are detects in blanks in all batches, establish the MRL as follows (otherwise proceed to 1C):

 □ Compare the lab’s reporting limit to the 90th percentile of field blanks (computed with non-detects set to ½ lab’s reporting limit). The higher value is the new MRL.

  ○ However, if we observe many detects in other types of blanks (e.g., matrix, solvent), we consider determining the MRL by comparing the lab’s reporting limit to the 90th percentile of ALL blanks (computed with non-detects set to ½ lab’s reporting limit). The higher value is the new MRL.

  ○ It can be helpful here to plot sample data with different possible MRLs to gain understanding of precisely what is being achieved by raising the MRL (i.e., are we successfully flagging data that we are not confident in and at the same time leaving data in which we have confidence unqualified?). See Additional file 4: Figure S9, for an example of this type of plot.

 *Note*: we use the 90th percentile of the blanks rather than using the maximum value or the mean because the 90th percentile is less sensitive to extreme values and can be estimated for data that are not normally distributed. However if the overall study is small (e.g., in our practice, when we have < 5 blanks), we set the MRL equal to the maximum blank mass.

1C. For each chemical detected in blanks, if detects in blanks are clustered in one or a few batches:

 □ If just one extremely problematic batch, consider dropping the sample data from that batch.

 □ If multiple field blanks were run in each batch, can consider determining MRL as above but on a batch-specific basis.

  ○ In this case, the way to proceed will very much be a judgment call. Spend time with the data considering various approaches.

 □ Data from reference material and duplicate samples can be helpful in deciding which data points should be qualified because they are “in the noise.”

2. After determining the MRL, we flag each sample result as follows:

 □ 0 flag = measurement reported by the lab as “non-detect”

 □ 0.5 flag = measurement falls below the MRL. These are considered “estimated detects”

 □ 1 flag = measurement falls above the MRL. These are considered “true detects”

Note that our data qualifier flags may differ from those used by others. For example, NHANES flags non-detects with a “1” and detects with a “0.”

3. Normalize MRL.

 □ If the MRL is determined on a mass basis but sample results are normalized by some factor, such as sample volume, we compute a sample-specific concentration-based MRL by dividing the mass-based MRL by the sample volume.

Reporting:

 □ We do not count estimated values (0.5 flags) as detects when reporting % > MRL. We do not use estimated detects to calculate summary statistics such as percentiles (see Table shell S2 in Additional file 2).

 □ In summary statistics, we identify any chemicals with greater than 50% estimated detects and add a footnote: “Imprecise quantification for more than 50% of detected values”.

 □ Graphical presentations should distinguish estimated from true detects (e.g., by plotting as different shapes, see Fig. 5).

 □ For reporting in tables, we use median sample volume across samples to convert mass-based MRL to a single concentration-based MRL for each chemical, if applicable.

 □ There are different approaches for incorporating estimated or 0.5 flagged values in statistical analyses, including performing analyses weighted by estimates of the measurement precision below the MRL, or using censored regression methods [23]. However any approach that incorporates estimated values is preferable to procedures that substitute with the DL, ½ DL, zero, or remove these values, a practice which can introduce bias [25].