What "not detected" actually means

My research background is microwave spectroscopy, which means I have spent a lot of time staring at the bottom of a signal trying to decide whether a feature is real. That is the whole job at the low end: not measuring a peak, but arguing about whether there is a peak. Everyone who does instrumental measurement for a living develops a particular allergy, and mine is to the phrase "not detected," because of how confidently it gets read as "not there."

Those are different statements. The distance between them is where a surprising amount of consumer and environmental confusion lives, and closing it does not require any chemistry. It requires reading one regulatory definition carefully.

The definition, and the word that gives it away

The United States has a legally binding definition of a detection limit, and it is worth quoting exactly. From 40 C.F.R. Part 136, Appendix B, the EPA's Method Detection Limit procedure:

"The method detection limit (MDL) is defined as the minimum measured concentration of a substance that can be reported with 99% confidence that the measured concentration is distinguishable from method blank results."

Read the end of that sentence, not the beginning. The comparison is not to zero. It is not to a health threshold, or a safety level, or the point where the substance stops mattering. The comparison is to method blank results: what the laboratory's own instrument reports when it is handed a sample that is supposed to contain nothing.

That is the entire trick, and once you see it the rest follows. A detection limit is the point at which a lab can distinguish your sample from its own background. The background belongs to the lab. So the limit belongs to the lab.

How the number gets made

The procedure in Appendix B is specific about this in a way that makes the point better than any argument. To establish an MDL, a laboratory analyzes a minimum of seven spiked samples and seven method blanks, prepared in at least three separate batches on three different calendar dates, and analyzed on three separate calendar dates.

It then computes two numbers. One, from the spiked replicates, uses the Student's t-value at the 99th percentile against their standard deviation. The other comes from the method blanks. And then the instruction that I find genuinely clarifying:

"Select the greater of MDLs or MDLb as the initial MDL."

Take the worse one. If your blanks are dirtier than your spikes are variable, your detection limit is set by your dirt.

Consider what is in that recipe. The lab's instrument. The lab's reagents. The lab's water. The lab's analysts. The particular three days it happened to run. None of those are properties of the substance being measured, and all of them move the number. Two competent, fully accredited laboratories running the same validated method on identical samples will publish different MDLs, and both will be correct, because they are answering a question about themselves.

Which means "not detected" is an abbreviation for something much longer: below the concentration at which this laboratory could distinguish this analyte from its own blanks, at 99% confidence, on the days it validated the method. Everything below that line is invisible, and invisible is not the same as absent.

Three different numbers wear the same coat

There is a second layer of confusion, and it comes from the fact that "the limit" is at least three limits.

The pharmaceutical world is unusually clear here. ICH Q2(R2), the harmonised guideline on validation of analytical procedures adopted on 1 November 2023, separates them explicitly. In the signal-to-noise approach it says: "The DL or QL are the minimum concentrations at which the analyte can be reliably detected or quantitated, respectively. A signal-to-noise ratio of 3:1 is generally considered acceptable for estimating the DL. For QL, a ratio of at least 10:1 is considered acceptable."

So there is a detection limit and, well above it, a quantitation limit. Where the guideline uses the standard deviation of the response and the slope of the calibration curve instead, the two are defined as DL = 3.3σ/S and QL = 10σ/S, roughly a factor of three apart.

The gap between them is a real region with a strange status. A result that lands there means the analyte is present, you can see it, and you cannot put a trustworthy number on it. "Detected but not quantified" is not a hedge or a dodge. It is a precise description of a measurement that has cleared one bar and not the other.

And then there is a third number, the reporting limit, which is the one that actually appears on the certificate you receive. ICH gestures at how far apart these can be when it allows a confirmatory validation step to be skipped "in cases where the QL is well below (e.g., approximately 10 times lower than) the reporting limit." An order of magnitude between what the method can quantify and what the lab reports is contemplated as routine.

So when a report says "not detected," ask which of the three numbers it is anchored to. Very often it is the reporting limit, which is the highest of them, chosen for defensibility rather than sensitivity. The lab is not being evasive. It is declining to defend numbers it does not want to defend, which is good practice. But it means the phrase covers more ground than it sounds like it does.

Where this stops being academic

The cleanest illustration I know is in American drinking water law, and it is unusually stark because the government publishes both numbers side by side.

Under the Safe Drinking Water Act, at 42 U.S.C. § 300g-1(b)(4)(A), a maximum contaminant level goal "shall be set at the level at which no known or anticipated adverse effects on the health of persons occur and which allows an adequate margin of safety." That is the pure health number, with no engineering in it.

Then subsection (b)(4)(B) says the enforceable standard, the maximum contaminant level, "shall specify a maximum contaminant level for such contaminant which is as close to the maximum contaminant level goal as is feasible." And (b)(4)(D) defines feasible in terms of "the best technology, treatment techniques and other means which the Administrator finds, after examination for efficacy under field conditions and not solely under laboratory conditions, are available (taking cost into consideration)."

Two numbers, deliberately. One says where harm stops. The other says how close we can actually get.

Now apply it. In the April 2024 PFAS National Primary Drinking Water Regulation, EPA set the maximum contaminant level goal for both PFOA and PFOS at zero, and the enforceable maximum contaminant level for each at 4.0 parts per trillion. EPA's current materials on the proposed compliance-extension rule restate that the MCLs "remain 4.0 parts per trillion (ppt) each," against an original compliance deadline of April 2029, with a May 18, 2026 proposal that would let systems request two additional years, to 2031, and would require short-term mitigation from systems at or above 12 ppt.

Sit with the pair. The health goal is zero. The enforceable limit is 4.0 parts per trillion. A water system can be fully compliant, be reported as meeting the standard, and still sit above the level the same agency identified as having an adequate margin of safety, because the enforceable number is a feasibility number and the goal is not.

And underneath both sits the measurement problem this whole piece is about. Four parts per trillion is roughly four drops in a thousand Olympic pools. Whether a given laboratory can see 3 ppt, or 6 ppt, is a question about that laboratory's blanks. If a lab's reporting limit for PFOA is above 4 ppt, its "not detected" is compatible with a result that would violate the standard. The words on the page are identical either way.

How to read a non-detect, in four questions

This is the practical residue, and it works on a water report, a soil report, a heavy-metals certificate of analysis on a supplement, or a residue screen.

First: what is the number? A non-detect without an accompanying limit is not a result, it is a mood. Any competent report states the limit next to it. If yours does not, that is the thing to ask for, and the answer should be a concentration with units.

Second: which limit is it? Detection, quantitation, or reporting. They can differ by ten times or more, and only the report can tell you which one is being used.

Third: is that limit below the number you care about? This is the one that matters and the one almost nobody asks. A non-detect is only reassuring if the limit sits comfortably under the threshold, the standard, or the label claim you are checking against. If the limit is above it, the test was not capable of answering your question, and it will still say "not detected" while failing to.

Fourth: is this comparable to the last one? Two reports on the same source from different labs, or from the same lab after a method change, are not directly comparable at the low end. A change from "detected at 5" to "not detected" can be a genuine improvement, or it can be the same water measured by an instrument with a higher floor.

The honest limits on this

Some things I am not saying. I am not saying non-detects are meaningless. They are informative, sometimes decisively so, and a low limit from a good lab is real evidence of absence in the only sense measurement ever provides. I am not saying laboratories are hiding anything: the definitions above are published, binding, and were written by the regulators themselves precisely so this is checkable.

I have also kept to what the sources say. The 40 C.F.R. Part 136 procedure is EPA's clean-water framework, and other domains use other formalisms, ISO and IUPAC among them, that partition detection and quantitation somewhat differently. I have used ICH Q2(R2) for the DL and QL distinction because it is explicit and current, not because pharmaceutical practice governs drinking water. And on PFAS I have quoted EPA's stated MCLG and MCL values and the statutory structure that produces two numbers; I have not attempted to reconstruct the internal reasoning behind the specific figure of 4.0, which is a detailed rulemaking record I did not open.

The signal

Every instrument has a floor, and the floor is made of the instrument's own noise. That is not a defect to be engineered away, it is what measurement is. You are always asking whether a signal stands above a background, and the background is yours.

So a detection limit is not a fact about a substance. It is a fact about a laboratory, on a set of days, with a set of reagents, running a defined procedure. "Not detected" reports the position of that floor. It says nothing whatsoever about what is underneath it.

The useful habit is small: whenever you see a non-detect, go find the number beside it, and ask whether that number is low enough to make the reassurance mean anything. Most of the time it is. The times it is not are exactly the times the phrase gets quoted hardest.