What a drinking-water average hides: censoring in the French monitoring record
Most results in a drinking-water monitoring programme are not measurements. They are statements that a substance was not detectable, published as a "less than" value alongside a numeric field that is usually zero. Read the numeric field and every trace determinand looks close to absent; the average is arithmetically correct and describes the laboratory rather than the water. This note sets out how much of the French record is censored, determinand by determinand, and what Hydrionis does about it.
How much of the record is a measurement
A regulatory drinking-water result is not always a number. When a substance is present below the concentration the laboratory can quantify, the result is published as a less than value — <0,5 — alongside a numeric field that is conventionally set to zero. The distinction is invisible if you read the numeric column, and it changes the answer completely.
| Determinand | Above detection | Departments | Year |
|---|---|---|---|
| PFOS | 11.8% | 93 | 2025 |
| Arsenic | 32.2% | 100 | 2025 |
| Lead | 35.7% | 54 | 2025 |
| Total pesticides | 43.2% | 100 | 2025 |
| Nitrate | 97.0% | 101 | 2025 |
Each figure is the median across French departments of the share of results that returned a number rather than a below-detection statement. Where that share is low, an average computed from the numeric column is dominated by zeroes that were never measured.
What Hydrionis does instead
Three rules, applied at ingestion rather than at display. Censored results are identified from the textual result, not the numeric field, so a zero-filled number cannot be mistaken for a measurement. They are excluded from every mean rather than substituted with zero or with half the detection limit, because substitution invents a measurement that was never made. And they are never counted as exceeding a regulatory limit, since being below detection is evidence of being below the limit.
Data behind this
Every figure above is computed from these datasets at build time rather than written into the text. If the underlying record changes, this page changes with it.
Citation
Hydrionis (2026), What a drinking-water average hides: censoring in the French monitoring record, version 1.0. Retrieved from https://hydrionis.com/research/what-a-drinking-water-average-hides/